init file sys

11 files changed, 1467 insertions, 50 deletions

diff --git a/backend/app/main.py b/backend/app/main.py
index 886bd9e..be30333 100644
--- a/backend/app/main.py
+++ b/backend/app/main.py
@@ -1,16 +1,20 @@
+import asyncio
+import tempfile
+import time
 from fastapi import FastAPI, HTTPException
 from fastapi.middleware.cors import CORSMiddleware
 from fastapi.responses import StreamingResponse, FileResponse
-from fastapi import UploadFile, File
+from fastapi import UploadFile, File, Form
 from pydantic import BaseModel
 from app.schemas import NodeRunRequest, NodeRunResponse, MergeStrategy, Role, Message, Context, LLMConfig, ModelProvider, ReasoningEffort
-from app.services.llm import llm_streamer, generate_title
+from app.services.llm import llm_streamer, generate_title, get_openai_client
 from dotenv import load_dotenv
 import os
 import json
 import shutil
 from typing import List, Literal, Optional
 from uuid import uuid4
+from google import genai
 
 load_dotenv()
 
@@ -28,6 +32,9 @@ app.add_middleware(
 DATA_ROOT = os.path.abspath(os.getenv("DATA_ROOT", os.path.join(os.getcwd(), "data")))
 DEFAULT_USER = "test"
 ARCHIVE_FILENAME = "archived_nodes.json"
+VALID_FILE_PROVIDERS = {"local", "openai", "google"}
+OPENAI_MAX_FILE_SIZE = 50 * 1024 * 1024  # 50MB limit per OpenAI docs
+OPENAI_DEFAULT_FILE_PURPOSE = os.getenv("OPENAI_FILE_PURPOSE", "user_data")
 
 def ensure_user_root(user: str) -> str:
     """
@@ -152,6 +159,10 @@ class FileMeta(BaseModel):
     created_at: float
     provider: Optional[str] = None
     provider_file_id: Optional[str] = None
+    openai_file_id: Optional[str] = None
+    openai_vector_store_id: Optional[str] = None
+    # Scopes for filtering: "project_path/node_id" composite keys
+    scopes: List[str] = []
 
 class FolderRequest(BaseModel):
     user: str = DEFAULT_USER
@@ -225,8 +236,44 @@ async def run_node_stream(request: NodeRunRequest):
 
     execution_context = Context(messages=final_messages)
 
+    tools: List[dict] = []
+    attachments: List[dict] = []
+
+    if request.config.provider == ModelProvider.OPENAI:
+        vs_ids, debug_refs, filters = await prepare_openai_vector_search(
+            user=DEFAULT_USER,
+            attached_ids=request.attached_file_ids,
+            scopes=request.scopes,
+            llm_config=request.config,
+        )
+        # Always enable file_search if vector store exists (even without explicit attachments)
+        # This allows nodes to access files attached in previous nodes of the trace
+        if not vs_ids:
+            # Try to get user's vector store anyway
+            try:
+                client = get_openai_client(request.config.api_key)
+                vs_id = await ensure_user_vector_store(DEFAULT_USER, client)
+                if vs_id:
+                    vs_ids = [vs_id]
+            except Exception as e:
+                print(f"[warn] Could not get vector store: {e}")
+        
+        if vs_ids:
+            tool_def = {"type": "file_search", "vector_store_ids": vs_ids}
+            if filters:
+                tool_def["filters"] = filters
+            tools.append(tool_def)
+        print(f"[openai file_search] vs_ids={vs_ids} refs={debug_refs} filters={filters}")
+    elif request.config.provider == ModelProvider.GOOGLE:
+        attachments = await prepare_attachments(
+            user=DEFAULT_USER,
+            target_provider=request.config.provider,
+            attached_ids=request.attached_file_ids,
+            llm_config=request.config,
+        )
+
     return StreamingResponse(
-        llm_streamer(execution_context, request.user_prompt, request.config),
+        llm_streamer(execution_context, request.user_prompt, request.config, attachments, tools),
         media_type="text/event-stream"
     )
 
@@ -396,6 +443,107 @@ def archived_path(user: str) -> str:
 def files_index_path(user: str) -> str:
     return os.path.join(files_root(user), "index.json")
 
+def user_vector_store_path(user: str) -> str:
+    return os.path.join(files_root(user), "vector_store.json")
+
+async def ensure_user_vector_store(user: str, client=None) -> str:
+    """
+    Ensure there is a vector store for the user (OpenAI).
+    Persist the id under data/<user>/files/vector_store.json.
+    """
+    path = user_vector_store_path(user)
+    if client is None:
+        client = get_openai_client()
+
+    # Try existing cached ID
+    if os.path.exists(path):
+        try:
+            with open(path, "r", encoding="utf-8") as f:
+                data = json.load(f)
+                vs_id_cached = data.get("id")
+                if vs_id_cached:
+                    try:
+                        await client.vector_stores.retrieve(vector_store_id=vs_id_cached)
+                        return vs_id_cached
+                    except Exception:
+                        # Possibly deleted; recreate below
+                        pass
+        except Exception:
+            pass
+
+    # create new
+    vs = await client.vector_stores.create(name=f"{user}-vs")
+    vs_id = getattr(vs, "id", None)
+    if not vs_id:
+        raise HTTPException(status_code=500, detail="Failed to create vector store")
+    os.makedirs(os.path.dirname(path), exist_ok=True)
+    with open(path, "w", encoding="utf-8") as f:
+        json.dump({"id": vs_id}, f)
+    return vs_id
+
+async def ensure_openai_file_and_index(user: str, meta: FileMeta, path: str, llm_config: Optional[LLMConfig] = None) -> tuple[str, str]:
+    """
+    Ensure the file is uploaded to OpenAI Files and added to the user's vector store.
+    Returns (openai_file_id, vector_store_id).
+    """
+    client = get_openai_client(llm_config.api_key if llm_config else None)
+    vs_id = await ensure_user_vector_store(user, client)
+
+    file_id = meta.openai_file_id or (meta.provider_file_id if meta.provider == "openai" else None)
+    if not file_id:
+        with open(path, "rb") as f:
+            content = f.read()
+        resp = await client.files.create(
+            file=(meta.name or "upload.bin", content),
+            purpose="assistants",
+        )
+        file_id = getattr(resp, "id", None)
+        if not file_id:
+            raise HTTPException(status_code=500, detail="OpenAI file upload returned no file_id")
+
+    await add_file_to_vector_store(vs_id, file_id, client=client)
+    return file_id, vs_id
+
+async def remove_file_from_vector_store(vs_id: str, file_id: str, client=None):
+    if not vs_id or not file_id:
+        return
+    if client is None:
+        client = get_openai_client()
+    try:
+        await client.vector_stores.files.delete(vector_store_id=vs_id, file_id=file_id)
+    except Exception as e:
+        print(f"[warn] remove_file_from_vector_store failed: {e}")
+
+async def add_file_to_vector_store(vs_id: str, file_id: str, client=None):
+    """
+    Add a file to vector store with file_id as attribute for filtering.
+    We use file_id as the attribute so we can filter by specific files at query time.
+    """
+    if client is None:
+        client = get_openai_client()
+    
+    # Use file_id as attribute for filtering
+    create_params = {
+        "vector_store_id": vs_id, 
+        "file_id": file_id,
+        "attributes": {"file_id": file_id}  # Enable filtering by file_id
+    }
+    
+    await client.vector_stores.files.create(**create_params)
+    # Poll until completed (limit capped at 100 per API spec)
+    for _ in range(20):
+        listing = await client.vector_stores.files.list(vector_store_id=vs_id, limit=100)
+        found = None
+        for item in getattr(listing, "data", []):
+            if getattr(item, "id", None) == file_id or getattr(item, "file_id", None) == file_id:
+                found = item
+                break
+        status = getattr(found, "status", None) if found else None
+        if status == "completed":
+            return
+        await asyncio.sleep(0.5)
+    # If not confirmed, still continue
+    return
 
 def load_files_index(user: str) -> List[FileMeta]:
     path = files_index_path(user)
@@ -412,6 +560,226 @@ def save_files_index(user: str, items: List[FileMeta]):
     with open(path, "w", encoding="utf-8") as f:
         json.dump([item.model_dump() for item in items], f, ensure_ascii=False, indent=2)
 
+
+async def prepare_attachments(
+    user: str,
+    target_provider: str,
+    attached_ids: List[str],
+    llm_config: LLMConfig,
+) -> list[dict]:
+    """
+    For each attached file ID:
+    - If already uploaded to the target provider, reuse provider_file_id/uri.
+    - Otherwise, upload with the original filename (required by OpenAI).
+    Returns a list of dicts describing attachment references for the provider.
+    """
+    if not attached_ids:
+        return []
+
+    items = load_files_index(user)
+    items_map = {item.id: item for item in items}
+    attachments: list[dict] = []
+
+    if isinstance(target_provider, ModelProvider):
+        provider_norm = target_provider.value.lower()
+    else:
+        provider_norm = str(target_provider).lower()
+
+    for fid in attached_ids:
+        meta = items_map.get(fid)
+        if not meta:
+            print(f"[warn] Attached file id not found, skipping: {fid}")
+            continue
+
+        path = os.path.join(files_root(user), fid)
+        if not os.path.exists(path):
+            raise HTTPException(status_code=404, detail=f"Attached file missing on disk: {meta.name}")
+
+        if provider_norm == ModelProvider.OPENAI or provider_norm == "openai":
+            # Reuse provider file id if available
+            if meta.provider == "openai" and meta.provider_file_id:
+                attachments.append({
+                    "provider": "openai",
+                    "file_id": meta.provider_file_id,
+                    "name": meta.name,
+                    "mime": meta.mime,
+                })
+                continue
+
+            # Upload to OpenAI with original filename
+            with open(path, "rb") as f:
+                content = f.read()
+            size = len(content)
+            if size > OPENAI_MAX_FILE_SIZE:
+                raise HTTPException(status_code=400, detail=f"File {meta.name} exceeds OpenAI 50MB limit")
+
+            try:
+                client = get_openai_client(llm_config.api_key)
+                resp = await client.files.create(
+                    file=(meta.name or "upload.bin", content),
+                    purpose=OPENAI_DEFAULT_FILE_PURPOSE,
+                )
+                openai_file_id = getattr(resp, "id", None)
+                if not openai_file_id:
+                    raise HTTPException(status_code=500, detail="OpenAI file upload returned no file_id")
+                attachments.append({
+                    "provider": "openai",
+                    "file_id": openai_file_id,
+                    "name": meta.name,
+                    "mime": meta.mime,
+                })
+            except HTTPException:
+                raise
+            except Exception as e:
+                raise HTTPException(status_code=500, detail=f"OpenAI upload failed: {str(e)}")
+
+        elif provider_norm == ModelProvider.GOOGLE or provider_norm == "google":
+            # Reuse uri/name if available and looks like a URI
+            if meta.provider == "google" and meta.provider_file_id and "://" in meta.provider_file_id:
+                attachments.append({
+                    "provider": "google",
+                    "uri": meta.provider_file_id,
+                    "name": meta.name,
+                    "mime": meta.mime,
+                })
+                continue
+
+            key = llm_config.api_key or os.getenv("GOOGLE_API_KEY")
+            if not key:
+                raise HTTPException(status_code=500, detail="Google API Key not found")
+            client = genai.Client(api_key=key)
+
+            tmp_path = None
+            try:
+                with open(path, "rb") as f:
+                    content = f.read()
+                with tempfile.NamedTemporaryFile(delete=False) as tmp:
+                    tmp.write(content)
+                    tmp_path = tmp.name
+
+                google_resp = await asyncio.to_thread(
+                    client.files.upload,
+                    file=tmp_path,
+                    config={"mimeType": meta.mime or "application/octet-stream"},
+                )
+                google_name = getattr(google_resp, "name", None)
+                google_uri = getattr(google_resp, "uri", None)
+
+                # Poll for ACTIVE and uri if missing
+                if google_name:
+                    for _ in range(10):
+                        try:
+                            info = await asyncio.to_thread(client.files.get, name=google_name)
+                            state = getattr(info, "state", None)
+                            google_uri = getattr(info, "uri", google_uri)
+                            if str(state).upper().endswith("ACTIVE") or state == "ACTIVE":
+                                break
+                            await asyncio.sleep(1)
+                        except Exception:
+                            await asyncio.sleep(1)
+                print(f"[google upload] name={google_name} uri={google_uri}")
+
+                uri = google_uri or google_name
+                if not uri:
+                    raise HTTPException(status_code=500, detail="Google upload returned no uri/name")
+                attachments.append({
+                    "provider": "google",
+                    "uri": uri,
+                    "name": meta.name,
+                    "mime": meta.mime,
+                })
+            except HTTPException:
+                raise
+            except Exception as e:
+                raise HTTPException(status_code=500, detail=f"Google upload failed: {str(e)}")
+            finally:
+                if tmp_path and os.path.exists(tmp_path):
+                    os.remove(tmp_path)
+
+        else:
+            raise HTTPException(status_code=400, detail=f"Unsupported provider for attachments: {target_provider}")
+
+    # Debug log
+    print(f"[attachments] provider={provider_norm} count={len(attachments)} detail={[{'name': a.get('name'), 'id': a.get('file_id', a.get('uri'))} for a in attachments]}")
+    return attachments
+
+
+async def prepare_openai_vector_search(
+    user: str,
+    attached_ids: List[str],
+    scopes: List[str],
+    llm_config: LLMConfig,
+) -> tuple[List[str], List[dict], Optional[dict]]:
+    """
+    Ensure all attached files are uploaded to OpenAI Files (purpose=assistants) and added to the user's vector store.
+    Returns (vector_store_ids, openai_file_refs_for_debug, filters).
+    
+    Filtering logic:
+    - If scopes provided: find files whose scopes intersect with requested scopes
+    - If only attached_ids: use those specific files
+    - Filters are constructed using file_id attribute in vector store
+    """
+    items = load_files_index(user)
+    items_map = {item.id: item for item in items}
+    
+    # Determine which files to include based on scopes or attached_ids
+    relevant_files: List[FileMeta] = []
+    
+    if scopes:
+        # Find files whose scopes intersect with requested scopes
+        for item in items:
+            if item.scopes and any(s in scopes for s in item.scopes):
+                relevant_files.append(item)
+        print(f"[file_search] scopes={scopes} matched_files={[f.name for f in relevant_files]}")
+    elif attached_ids:
+        # Fallback: use explicitly attached files
+        for fid in attached_ids:
+            meta = items_map.get(fid)
+            if meta:
+                relevant_files.append(meta)
+    
+    if not relevant_files:
+        return [], [], None
+
+    changed = False
+    vs_ids: List[str] = []
+    debug_refs: List[dict] = []
+    file_ids_for_filter: List[str] = []
+
+    for meta in relevant_files:
+        path = os.path.join(files_root(user), meta.id)
+        if not os.path.exists(path):
+            print(f"[warn] Attached file missing on disk, skipping: {meta.id}")
+            continue
+        # Enforce 50MB OpenAI limit
+        file_size = os.path.getsize(path)
+        if file_size > OPENAI_MAX_FILE_SIZE:
+            print(f"[warn] File {meta.name} exceeds OpenAI 50MB limit, skipping")
+            continue
+
+        openai_file_id, vs_id = await ensure_openai_file_and_index(user, meta, path, llm_config)
+        if meta.openai_file_id != openai_file_id or meta.openai_vector_store_id != vs_id:
+            meta.openai_file_id = openai_file_id
+            meta.openai_vector_store_id = vs_id
+            changed = True
+        vs_ids.append(vs_id)
+        debug_refs.append({"name": meta.name, "file_id": openai_file_id, "vs_id": vs_id})
+        if openai_file_id:
+            file_ids_for_filter.append(openai_file_id)
+
+    if changed:
+        save_files_index(user, list(items_map.values()))
+
+    # deduplicate
+    vs_ids_unique = list({vid for vid in vs_ids if vid})
+    
+    # Build filters to only search relevant files
+    filters = None
+    if file_ids_for_filter:
+        filters = {"type": "in", "key": "file_id", "value": file_ids_for_filter}
+    
+    return vs_ids_unique, debug_refs, filters
+
 # -------------------------------------------------
 
 @app.get("/api/projects/archived")
@@ -449,27 +817,124 @@ def list_files(user: str = DEFAULT_USER):
 
 
 @app.post("/api/files/upload")
-async def upload_file(user: str = DEFAULT_USER, file: UploadFile = File(...)):
+async def upload_file(
+    user: str = DEFAULT_USER,
+    file: UploadFile = File(...),
+    provider: str = Form("local"),
+    purpose: Optional[str] = Form(None),
+):
     migrate_legacy_layout(user)
     items = load_files_index(user)
     file_id = str(uuid4())
     dest_root = files_root(user)
     dest_path = os.path.join(dest_root, file_id)
+    file_name = file.filename or "upload.bin"
+    provider_normalized = (provider or "local").lower()
+    if provider_normalized not in VALID_FILE_PROVIDERS:
+        raise HTTPException(status_code=400, detail="Unsupported provider")
+
     try:
         content = await file.read()
-        size = len(content)
+    except Exception as e:
+        raise HTTPException(status_code=500, detail=str(e))
+
+    size = len(content)
+    if provider_normalized == "openai" and size > OPENAI_MAX_FILE_SIZE:
+        raise HTTPException(status_code=400, detail="OpenAI provider limit: max 50MB per file")
+
+    provider_file_id: Optional[str] = None
+    provider_created_at: Optional[float] = None
+
+    if provider_normalized == "openai":
+        try:
+            client = get_openai_client()
+            upload_purpose = purpose or OPENAI_DEFAULT_FILE_PURPOSE
+            resp = await client.files.create(
+                file=(file_name, content),
+                purpose=upload_purpose,
+            )
+            provider_file_id = getattr(resp, "id", None)
+            provider_created_at = getattr(resp, "created_at", None)
+        except Exception as e:
+            raise HTTPException(status_code=500, detail=f"OpenAI upload failed: {str(e)}")
+    elif provider_normalized == "google":
+        try:
+            key = os.getenv("GOOGLE_API_KEY")
+            if not key:
+                raise HTTPException(status_code=500, detail="Google API Key not found")
+            client = genai.Client(api_key=key)
+            # The Google GenAI SDK upload is synchronous; run in thread to avoid blocking the event loop.
+            tmp_path = None
+            try:
+                with tempfile.NamedTemporaryFile(delete=False) as tmp:
+                    tmp.write(content)
+                    tmp_path = tmp.name
+                google_resp = await asyncio.to_thread(
+                    client.files.upload,
+                    file=tmp_path,
+                    config={"mimeType": file.content_type or "application/octet-stream"},
+                )
+                google_name = getattr(google_resp, "name", None)
+                google_uri = getattr(google_resp, "uri", None)
+
+                # Poll for ACTIVE and uri if missing
+                if google_name:
+                    for _ in range(10):
+                        try:
+                            info = await asyncio.to_thread(client.files.get, name=google_name)
+                            state = getattr(info, "state", None)
+                            google_uri = getattr(info, "uri", google_uri)
+                            if str(state).upper().endswith("ACTIVE") or state == "ACTIVE":
+                                break
+                            await asyncio.sleep(1)
+                        except Exception:
+                            await asyncio.sleep(1)
+
+                provider_file_id = google_uri or google_name
+            finally:
+                if tmp_path and os.path.exists(tmp_path):
+                    os.remove(tmp_path)
+
+            provider_created_at = time.time()
+        except HTTPException:
+            raise
+        except Exception as e:
+            raise HTTPException(status_code=500, detail=f"Google upload failed: {str(e)}")
+
+    try:
+        os.makedirs(dest_root, exist_ok=True)
         with open(dest_path, "wb") as f:
             f.write(content)
     except Exception as e:
         raise HTTPException(status_code=500, detail=str(e))
 
+    created_at = provider_created_at or os.path.getmtime(dest_path)
+
     meta = FileMeta(
         id=file_id,
-        name=file.filename,
+        name=file_name,
         size=size,
         mime=file.content_type or "application/octet-stream",
-        created_at=os.path.getmtime(dest_path),
+        created_at=created_at,
+        provider=provider_normalized if provider_normalized != "local" else None,
+        provider_file_id=provider_file_id,
+        openai_file_id=None,
+        openai_vector_store_id=None,
     )
+
+    # Always try to index into OpenAI vector store (if <=50MB)
+    if size <= OPENAI_MAX_FILE_SIZE:
+        try:
+            openai_file_id, vs_id = await ensure_openai_file_and_index(user, meta, dest_path, None)
+            meta.openai_file_id = openai_file_id
+            meta.openai_vector_store_id = vs_id
+            if provider_normalized == "openai" and not meta.provider_file_id:
+                meta.provider_file_id = openai_file_id
+        except Exception as e:
+            print(f"[warn] OpenAI indexing failed for {file_name}: {e}")
+    else:
+        print(f"[warn] Skipping OpenAI indexing for {file_name}: exceeds 50MB")
+
     items.append(meta)
     save_files_index(user, items)
     return {"file": meta}
@@ -489,15 +954,87 @@ def download_file(user: str = DEFAULT_USER, file_id: str = ""):
 
 
 @app.post("/api/files/delete")
-def delete_file(user: str = DEFAULT_USER, file_id: str = ""):
+async def delete_file(user: str = DEFAULT_USER, file_id: str = ""):
     migrate_legacy_layout(user)
     items = load_files_index(user)
     meta = next((i for i in items if i.id == file_id), None)
     if not meta:
         raise HTTPException(status_code=404, detail="file not found")
+
+    # Remove from vector store and OpenAI Files if present
+    if meta.openai_vector_store_id and meta.openai_file_id:
+        await remove_file_from_vector_store(meta.openai_vector_store_id, meta.openai_file_id)
+    if meta.provider == "openai" and meta.provider_file_id:
+        try:
+            client = get_openai_client()
+            await client.files.delete(meta.provider_file_id)
+        except Exception as e:
+            raise HTTPException(status_code=500, detail=f"OpenAI delete failed: {str(e)}")
+    if meta.provider == "google" and meta.provider_file_id:
+        try:
+            key = os.getenv("GOOGLE_API_KEY")
+            if not key:
+                raise HTTPException(status_code=500, detail="Google API Key not found")
+            client = genai.Client(api_key=key)
+            await asyncio.to_thread(client.files.delete, meta.provider_file_id)
+        except HTTPException:
+            raise
+        except Exception as e:
+            raise HTTPException(status_code=500, detail=f"Google delete failed: {str(e)}")
+
     path = os.path.join(files_root(user), file_id)
     if os.path.exists(path):
         os.remove(path)
     items = [i for i in items if i.id != file_id]
     save_files_index(user, items)
     return {"ok": True}
+
+
+class AddScopeRequest(BaseModel):
+    user: str = DEFAULT_USER
+    file_id: str
+    scope: str  # "project_path/node_id" composite key
+
+
+@app.post("/api/files/add_scope")
+def add_file_scope(request: AddScopeRequest):
+    """
+    Add a scope to a file's scopes list.
+    Called when user attaches a file to a node.
+    """
+    migrate_legacy_layout(request.user)
+    items = load_files_index(request.user)
+    meta = next((i for i in items if i.id == request.file_id), None)
+    if not meta:
+        raise HTTPException(status_code=404, detail="file not found")
+    
+    if request.scope not in meta.scopes:
+        meta.scopes.append(request.scope)
+        save_files_index(request.user, items)
+    
+    return {"file": meta.model_dump()}
+
+
+class RemoveScopeRequest(BaseModel):
+    user: str = DEFAULT_USER
+    file_id: str
+    scope: str
+
+
+@app.post("/api/files/remove_scope")
+def remove_file_scope(request: RemoveScopeRequest):
+    """
+    Remove a scope from a file's scopes list.
+    Called when user detaches a file from a node.
+    """
+    migrate_legacy_layout(request.user)
+    items = load_files_index(request.user)
+    meta = next((i for i in items if i.id == request.file_id), None)
+    if not meta:
+        raise HTTPException(status_code=404, detail="file not found")
+    
+    if request.scope in meta.scopes:
+        meta.scopes.remove(request.scope)
+        save_files_index(request.user, items)
+    
+    return {"file": meta.model_dump()}
diff --git a/backend/app/schemas.py b/backend/app/schemas.py
index bd8ebe7..54c0560 100644
--- a/backend/app/schemas.py
+++ b/backend/app/schemas.py
@@ -49,6 +49,10 @@ class NodeRunRequest(BaseModel):
     user_prompt: str
     config: LLMConfig
     merge_strategy: MergeStrategy = MergeStrategy.SMART
+    attached_file_ids: List[str] = Field(default_factory=list)
+    # Scopes for file_search filtering: ["project_path/node_id", ...]
+    # Contains all project/node combinations in the current trace
+    scopes: List[str] = Field(default_factory=list)
     
 class NodeRunResponse(BaseModel):
     node_id: str
diff --git a/backend/app/services/llm.py b/backend/app/services/llm.py
index b372f9e..96b0514 100644
--- a/backend/app/services/llm.py
+++ b/backend/app/services/llm.py
@@ -1,5 +1,5 @@
 import os
-from typing import AsyncGenerator
+from typing import AsyncGenerator, List, Dict, Any, Optional
 import openai
 import google.generativeai as genai
 from app.schemas import LLMConfig, Message, Role, Context
@@ -23,8 +23,15 @@ def configure_google(api_key: str = None):
         raise ValueError("Google API Key not found")
     genai.configure(api_key=key)
 
-async def stream_openai(messages: list[Message], config: LLMConfig) -> AsyncGenerator[str, None]:
+async def stream_openai(
+    messages: list[Message],
+    config: LLMConfig,
+    attachments: Optional[List[Dict[str, Any]]] = None,
+    tools: Optional[List[Dict[str, Any]]] = None,
+) -> AsyncGenerator[str, None]:
     client = get_openai_client(config.api_key)
+    attachments = attachments or []
+    tools = tools or []
     
     # Convert internal Message schema to OpenAI format
     openai_messages = []
@@ -38,17 +45,23 @@ async def stream_openai(messages: list[Message], config: LLMConfig) -> AsyncGene
     responses_only_models = ['gpt-5-pro']
     
     # Models that CAN use Responses API (and thus support web_search tool)
+    model_lower = config.model_name.lower()
     responses_capable_models = [
-        'gpt-5', 'gpt-5-chat-latest', 'gpt-5-mini', 'gpt-5-nano', 
-        'gpt-5-pro', 'gpt-5.1', 'gpt-5.1-chat-latest', 'o3'
+        'gpt-5', 'gpt-5-chat-latest', 'gpt-5-mini', 'gpt-5-nano',
+        'gpt-5-pro', 'gpt-5.1', 'gpt-5.1-chat-latest', 'o3',
+        'o1', 'o1-preview', 'o1-mini',
+        'gpt-4o', 'gpt-4o-mini', 'gpt-4o-realtime', 'gpt-4o-mini-tts'
     ]
     
     # Use Responses API if:
     # 1. Model ONLY supports Responses API, OR
     # 2. User wants web search AND model is capable of Responses API
+    # 3. Attachments are present (Responses supports input_file)
     use_responses_api = (
         config.model_name in responses_only_models or 
-        (config.enable_google_search and config.model_name in responses_capable_models)
+        (config.enable_google_search and (config.model_name in responses_capable_models or model_lower.startswith("gpt-4o"))) or
+        (attachments and (config.model_name in responses_capable_models or model_lower.startswith("gpt-4o"))) or
+        (tools)
     )
 
     if use_responses_api:
@@ -56,25 +69,50 @@ async def stream_openai(messages: list[Message], config: LLMConfig) -> AsyncGene
         # yield f"[Debug: Config Search={config.enable_google_search}, Model={config.model_name}]\n" 
 
         # Use new client.responses.create API with Polling Strategy
-        # Convert messages to Responses API format (same as Chat Completions)
-        # Responses API accepts input as array of message objects
-        
-        # Filter out system messages (use instructions instead) and format for Responses API
+        # Build Responses API input
         input_messages = []
         for msg in openai_messages:
-            if msg['role'] != 'system':  # System prompt goes to instructions
-                input_messages.append({
-                    "role": msg['role'],
-                    "content": msg['content']
+            if msg['role'] == 'system':
+                continue  # goes to instructions
+            # User messages use input_text, assistant messages use output_text
+            content_type = "input_text" if msg['role'] == 'user' else "output_text"
+            input_messages.append({
+                "role": msg['role'],
+                "content": [
+                    {
+                        "type": content_type,
+                        "text": msg['content']
+                    }
+                ]
+            })
+
+        # Append attachments as separate user message (files only)
+        file_parts = []
+        for att in attachments:
+            if att.get("provider") == "openai" and att.get("file_id"):
+                file_parts.append({
+                    "type": "input_file",
+                    "file_id": att["file_id"]
                 })
+        if file_parts:
+            input_messages.append({
+                "role": "user",
+                "content": file_parts
+            })
         
         resp_params = {
             "model": config.model_name,
             "input": input_messages,  # Full conversation history
-            "stream": False,    # Disable stream to get immediate ID
-            "background": True, # Enable background mode for async execution
-            "store": True
+            "stream": False,    # Get full output in one call
+            "background": False,
+            "store": True,
+            "tool_choice": "auto",
         }
+        if tools:
+            resp_params["tools"] = tools
+            resp_params["tool_choice"] = "auto"
+        # Optional: include results for debugging / citations
+        resp_params["include"] = ["file_search_call.results"]
         
         # Add reasoning effort (not supported by chat-latest models)
         models_without_effort = ['gpt-5-chat-latest', 'gpt-5.1-chat-latest']
@@ -82,28 +120,40 @@ async def stream_openai(messages: list[Message], config: LLMConfig) -> AsyncGene
             resp_params["reasoning"] = {"effort": config.reasoning_effort.value}
         
         # Enable Web Search if requested (Reusing enable_google_search flag as generic web_search flag)
+        # IMPORTANT: Append to existing tools instead of overwriting
         if config.enable_google_search:
-            resp_params["tools"] = [{"type": "web_search"}]
+            if resp_params.get("tools"):
+                resp_params["tools"].append({"type": "web_search"})
+            else:
+                resp_params["tools"] = [{"type": "web_search"}]
             resp_params["tool_choice"] = "auto"
-            # Debugging tool injection
-            # yield "[Debug: Web Search Tool Injected]" # Uncomment to debug
         
         if config.system_prompt:
             resp_params["instructions"] = config.system_prompt
+
+        # Debug: print final tools being sent
+        print(f"[responses debug] final tools: {resp_params.get('tools')}")
             
-        # 1. Create Response (Async/Background)
-        # This returns a Response object immediately with status 'queued' or 'in_progress'
+        # 1. Create Response (non-background)
         initial_resp = await client.responses.create(**resp_params)
         response_id = initial_resp.id
-        
+
         # 2. Poll for Completion
         import asyncio
-        # Poll for up to 10 minutes
-        for _ in range(300): 
+        for _ in range(300):
             final_resp = await client.responses.retrieve(response_id)
-            
+
             if final_resp.status == 'completed':
-                # Parse final response object
+                # Debug: log outputs and tool calls
+                try:
+                    outs = getattr(final_resp, "output", [])
+                    print(f"[responses debug] output items: {[getattr(o, 'type', None) for o in outs]}")
+                    for o in outs:
+                        if getattr(o, "type", None) == "file_search_call":
+                            print(f"[responses debug] file_search_call: {o}")
+                except Exception as e:
+                    print(f"[responses debug] failed to inspect output: {e}")
+
                 found_content = False
                 if hasattr(final_resp, 'output'):
                     for out in final_resp.output:
@@ -128,13 +178,16 @@ async def stream_openai(messages: list[Message], config: LLMConfig) -> AsyncGene
                 yield f"\n[Error: Response generation {final_resp.status}: {error_msg}]"
                 return
             
-            # Still in_progress
             await asyncio.sleep(2)
             
         yield "\n[Error: Polling timed out]"
         return
 
-    # Standard Chat Completions API
+    # Standard Chat Completions API (attachments not supported here)
+    if attachments:
+        yield "[Error] Attachments are only supported for Responses API-capable models."
+        return
+
     # Prepare parameters
     req_params = {
         "model": config.model_name,
@@ -175,7 +228,8 @@ async def stream_openai(messages: list[Message], config: LLMConfig) -> AsyncGene
             elif getattr(delta, 'refusal', None):
                 yield f"[Refusal: {delta.refusal}]"
 
-async def stream_google(messages: list[Message], config: LLMConfig) -> AsyncGenerator[str, None]:
+async def stream_google(messages: list[Message], config: LLMConfig, attachments: List[Dict[str, Any]] | None = None) -> AsyncGenerator[str, None]:
+    attachments = attachments or []
     # Use new Google GenAI SDK (google-genai)
     from google import genai
     from google.genai import types
@@ -200,6 +254,34 @@ async def stream_google(messages: list[Message], config: LLMConfig) -> AsyncGene
         tools=tools
     )
     
+    # If attachments present, send as a single generate_content call (non-streaming)
+    if attachments:
+        parts = []
+        for att in attachments:
+            uri = att.get("uri")
+            mime = att.get("mime") or "application/octet-stream"
+            if uri:
+                try:
+                    parts.append(types.Part.from_uri(uri, mime_type=mime))
+                except Exception:
+                    parts.append(types.Part(text=f"[file attached: {uri}]"))
+        for msg in messages:
+            parts.append(types.Part(text=msg.content))
+        print(f"[gemini] sending attachments: {[att.get('uri') for att in attachments]}")
+        try:
+            response = await client.aio.models.generate_content(
+                model=config.model_name,
+                contents=[types.Content(role="user", parts=parts)],
+                config=gen_config
+            )
+            if response and getattr(response, "text", None):
+                yield response.text
+            else:
+                yield "[Error] Gemini response returned no text."
+        except Exception as e:
+            yield f"[Error] Gemini call failed: {str(e)}"
+        return
+
     # Prepare History
     # Extract last message as the prompt
     prompt_msg = "..."
@@ -223,13 +305,6 @@ async def stream_google(messages: list[Message], config: LLMConfig) -> AsyncGene
         config=gen_config
     )
     
-    # Streaming call
-    # In google-genai SDK, streaming is usually via send_message_stream
-    
-    # Check if send_message_stream exists, otherwise use send_message with stream=True (but error says no)
-    # Let's assume send_message_stream is the way.
-    
-    # Note: chat_session.send_message_stream returns an AsyncIterator (or a coroutine returning one)
     response_stream = await chat_session.send_message_stream(prompt_msg)
     
     async for chunk in response_stream:
@@ -237,7 +312,13 @@ async def stream_google(messages: list[Message], config: LLMConfig) -> AsyncGene
         if chunk.text:
             yield chunk.text
 
-async def llm_streamer(context: Context, user_prompt: str, config: LLMConfig) -> AsyncGenerator[str, None]:
+async def llm_streamer(
+    context: Context,
+    user_prompt: str,
+    config: LLMConfig,
+    attachments: List[Dict[str, Any]] | None = None,
+    tools: List[Dict[str, Any]] | None = None,
+) -> AsyncGenerator[str, None]:
     # 1. Merge Context + New User Prompt
     # We create a temporary list of messages for this inference
     messages_to_send = context.messages.copy()
@@ -253,10 +334,10 @@ async def llm_streamer(context: Context, user_prompt: str, config: LLMConfig) ->
     # 2. Call Provider
     try:
         if config.provider == "openai":
-            async for chunk in stream_openai(messages_to_send, config):
+            async for chunk in stream_openai(messages_to_send, config, attachments, tools):
                 yield chunk
         elif config.provider == "google":
-            async for chunk in stream_google(messages_to_send, config):
+            async for chunk in stream_google(messages_to_send, config, attachments):
                 yield chunk
         else:
             yield f"Error: Unsupported provider {config.provider}"
diff --git a/backend/data/test/files/1d344051-3b8e-41bc-8c57-15710383bd42 b/backend/data/test/files/1d344051-3b8e-41bc-8c57-15710383bd42
new file mode 100644
index 0000000..747674f
--- /dev/null
+++ b/backend/data/test/files/1d344051-3b8e-41bc-8c57-15710383bd42
diff --git a/backend/data/test/files/58252c05-718e-4eea-a23d-56a64d03b4fc b/backend/data/test/files/23f72737-1d5f-4cab-b68e-8c9010a8fdc4
index a837816..a837816 100644
--- a/backend/data/test/files/58252c05-718e-4eea-a23d-56a64d03b4fc
+++ b/backend/data/test/files/23f72737-1d5f-4cab-b68e-8c9010a8fdc4
diff --git a/backend/data/test/files/2d43e6a5-2823-4564-8beb-fef2721056f3 b/backend/data/test/files/2d43e6a5-2823-4564-8beb-fef2721056f3
new file mode 100644
index 0000000..c522464
--- /dev/null
+++ b/backend/data/test/files/2d43e6a5-2823-4564-8beb-fef2721056f3
diff --git a/backend/data/test/files/609ece1b-0894-4904-b363-d6e1c7a90be8 b/backend/data/test/files/609ece1b-0894-4904-b363-d6e1c7a90be8
new file mode 100644
index 0000000..f2d6870
--- /dev/null
+++ b/backend/data/test/files/609ece1b-0894-4904-b363-d6e1c7a90be8
diff --git a/backend/data/test/files/dcf69ca5-2efa-4b8a-acec-3864a803f073 b/backend/data/test/files/dcf69ca5-2efa-4b8a-acec-3864a803f073
new file mode 100644
index 0000000..0b9b64e
--- /dev/null
+++ b/backend/data/test/files/dcf69ca5-2efa-4b8a-acec-3864a803f073
diff --git a/backend/data/test/files/index.json b/backend/data/test/files/index.json
index d9e559f..537cbe1 100644
--- a/backend/data/test/files/index.json
+++ b/backend/data/test/files/index.json
@@ -1,11 +1,75 @@
 [
   {
-    "id": "58252c05-718e-4eea-a23d-56a64d03b4fc",
+    "id": "dcf69ca5-2efa-4b8a-acec-3864a803f073",
+    "name": "542proposal.pdf",
+    "size": 275191,
+    "mime": "application/pdf",
+    "created_at": 1765347496.7856553,
+    "provider": null,
+    "provider_file_id": null,
+    "openai_file_id": "file-UTunaRvEEufAK3vMPtC4oh",
+    "openai_vector_store_id": "vs_693910aa910081919231c2c6b6c69e1d",
+    "scopes": [
+      "untitled/node_1765410411339",
+      "untitled.json/node_1765320756261",
+      "untitled (2).json/node_1765320756261"
+    ]
+  },
+  {
+    "id": "2d43e6a5-2823-4564-8beb-fef2721056f3",
+    "name": "ICLR_New (2).pdf",
+    "size": 4396443,
+    "mime": "application/pdf",
+    "created_at": 1765411275.0,
+    "provider": "openai",
+    "provider_file_id": "file-KF3avxznCxcDXnZxEEUhTo",
+    "openai_file_id": "file-KF3avxznCxcDXnZxEEUhTo",
+    "openai_vector_store_id": "vs_693910aa910081919231c2c6b6c69e1d",
+    "scopes": [
+      "untitled/node_1765410434473",
+      "untitled (2).json/node_1765411626468",
+      "untitled (2).json/node_1765412541501",
+      "untitled (2).json/node_1765413048314"
+    ]
+  },
+  {
+    "id": "23f72737-1d5f-4cab-b68e-8c9010a8fdc4",
     "name": "survey_on_survey (3).pdf",
     "size": 263914,
     "mime": "application/pdf",
-    "created_at": 1765324835.7883003,
+    "created_at": 1765415182.5394356,
+    "provider": null,
+    "provider_file_id": null,
+    "openai_file_id": "file-M1aQQuGPqBYPg48Wf7JfNo",
+    "openai_vector_store_id": "vs_693910aa910081919231c2c6b6c69e1d",
+    "scopes": []
+  },
+  {
+    "id": "609ece1b-0894-4904-b363-d6e1c7a90be8",
+    "name": "CV.pdf",
+    "size": 30337,
+    "mime": "application/pdf",
+    "created_at": 1765415665.079434,
+    "provider": null,
+    "provider_file_id": null,
+    "openai_file_id": "file-HB9LZHueJAaj9tNfx5QxvR",
+    "openai_vector_store_id": "vs_693910aa910081919231c2c6b6c69e1d",
+    "scopes": [
+      "untitled (2).json/quick_chat_temp"
+    ]
+  },
+  {
+    "id": "1d344051-3b8e-41bc-8c57-15710383bd42",
+    "name": "hw4.pdf",
+    "size": 249853,
+    "mime": "application/pdf",
+    "created_at": 1765416241.6634274,
     "provider": null,
-    "provider_file_id": null
+    "provider_file_id": null,
+    "openai_file_id": "file-UiqdRSQiyhcp4bNbAcuy7j",
+    "openai_vector_store_id": "vs_693910aa910081919231c2c6b6c69e1d",
+    "scopes": [
+      "untitled (2).json/quick_chat_temp"
+    ]
   }
 ]
\ No newline at end of file
diff --git a/backend/data/test/files/vector_store.json b/backend/data/test/files/vector_store.json
new file mode 100644
index 0000000..9b9d8a8
--- /dev/null
+++ b/backend/data/test/files/vector_store.json
@@ -0,0 +1 @@
+{"id": "vs_693910aa910081919231c2c6b6c69e1d"}
\ No newline at end of file
diff --git a/backend/data/test/projects/untitled (2).json b/backend/data/test/projects/untitled (2).json
new file mode 100644
index 0000000..2751b9c
--- /dev/null
+++ b/backend/data/test/projects/untitled (2).json
@@ -0,0 +1,730 @@
+{
+  "version": 1,
+  "nodes": [
+    {
+      "id": "node_1765320756261",
+      "type": "llmNode",
+      "position": {
+        "x": 100,
+        "y": 100
+      },
+      "data": {
+        "label": "PDF File Count",
+        "model": "gpt-5.1",
+        "temperature": 0.7,
+        "systemPrompt": "",
+        "userPrompt": "testing search_files. Please search for pdf files in this conversation and count the number.",
+        "mergeStrategy": "smart",
+        "reasoningEffort": "medium",
+        "messages": [
+          {
+            "id": "msg_1765413079192_u",
+            "role": "user",
+            "content": "testing search_files. Please search for pdf files in this conversation and count the number."
+          },
+          {
+            "id": "msg_1765413079192_a",
+            "role": "assistant",
+            "content": "I found **1 PDF file** in this conversation: `542proposal.pdf`. "
+          }
+        ],
+        "traces": [],
+        "outgoingTraces": [
+          {
+            "id": "fork-node_1765320756261-1765413083937",
+            "sourceNodeId": "node_1765320756261",
+            "color": "hsl(163, 70%, 60%)",
+            "messages": [
+              {
+                "id": "node_1765320756261-user",
+                "role": "user",
+                "content": "testing search_files. Please search for pdf files in this conversation and count the number."
+              },
+              {
+                "id": "node_1765320756261-assistant",
+                "role": "assistant",
+                "content": "I found **1 PDF file** in this conversation: `542proposal.pdf`. "
+              }
+            ]
+          }
+        ],
+        "forkedTraces": [
+          {
+            "id": "fork-node_1765320756261-1765413083937",
+            "sourceNodeId": "node_1765320756261",
+            "color": "hsl(163, 70%, 60%)",
+            "messages": [
+              {
+                "id": "node_1765320756261-u",
+                "role": "user",
+                "content": "testing search_files. Please search for pdf files in this conversation and count the number."
+              },
+              {
+                "id": "node_1765320756261-a",
+                "role": "assistant",
+                "content": "I found **1 PDF file** in this conversation: `542proposal.pdf`. "
+              }
+            ]
+          }
+        ],
+        "mergedTraces": [],
+        "response": "I found **1 PDF file** in this conversation: `542proposal.pdf`. ",
+        "status": "success",
+        "inputs": 1,
+        "activeTraceIds": [],
+        "attachments": [
+          {
+            "fileId": "8a29c29f-fc55-4197-ba61-ac6fa1b5d628",
+            "name": "507hw_.pdf",
+            "providerFileId": "file-9ptprToszH4K38Mg3Qu5B6",
+            "expiresAt": null
+          }
+        ],
+        "attachedFileIds": [
+          "dcf69ca5-2efa-4b8a-acec-3864a803f073"
+        ],
+        "querySentAt": 1765413043481,
+        "responseReceivedAt": 1765413079192
+      },
+      "width": 200,
+      "height": 64,
+      "selected": false,
+      "dragging": false,
+      "positionAbsolute": {
+        "x": 109.22690128581917,
+        "y": 455.2354603148382
+      }
+    },
+    {
+      "id": "node_1765413048314",
+      "type": "llmNode",
+      "position": {
+        "x": 450,
+        "y": 100
+      },
+      "data": {
+        "label": "Pdf Files",
+        "model": "gpt-5.1",
+        "temperature": 0.7,
+        "systemPrompt": "",
+        "userPrompt": "what about now?",
+        "mergeStrategy": "smart",
+        "reasoningEffort": "medium",
+        "messages": [
+          {
+            "id": "node_1765320756261-user",
+            "role": "user",
+            "content": "testing search_files. Please search for pdf files in this conversation and count the number."
+          },
+          {
+            "id": "node_1765320756261-assistant",
+            "role": "assistant",
+            "content": "I found **1 PDF file** in this conversation: `542proposal.pdf`. "
+          },
+          {
+            "id": "msg_1765415155872_u",
+            "role": "user",
+            "content": "what about now?"
+          },
+          {
+            "id": "msg_1765415155872_a",
+            "role": "assistant",
+            "content": "There are now **2 PDF files** available in this conversation:\n\n1. `ICLR_New (2).pdf`   \n2. `542proposal.pdf` "
+          }
+        ],
+        "traces": [
+          {
+            "id": "fork-node_1765320756261-1765413083937",
+            "sourceNodeId": "node_1765320756261",
+            "color": "hsl(163, 70%, 60%)",
+            "messages": [
+              {
+                "id": "node_1765320756261-user",
+                "role": "user",
+                "content": "testing search_files. Please search for pdf files in this conversation and count the number."
+              },
+              {
+                "id": "node_1765320756261-assistant",
+                "role": "assistant",
+                "content": "I found **1 PDF file** in this conversation: `542proposal.pdf`. "
+              }
+            ]
+          }
+        ],
+        "outgoingTraces": [
+          {
+            "id": "fork-node_1765320756261-1765413083937",
+            "sourceNodeId": "node_1765320756261",
+            "color": "hsl(163, 70%, 60%)",
+            "messages": [
+              {
+                "id": "node_1765320756261-user",
+                "role": "user",
+                "content": "testing search_files. Please search for pdf files in this conversation and count the number."
+              },
+              {
+                "id": "node_1765320756261-assistant",
+                "role": "assistant",
+                "content": "I found **1 PDF file** in this conversation: `542proposal.pdf`. "
+              },
+              {
+                "id": "node_1765413048314-user",
+                "role": "user",
+                "content": "what about now?"
+              },
+              {
+                "id": "node_1765413048314-assistant",
+                "role": "assistant",
+                "content": "There are now **2 PDF files** available in this conversation:\n\n1. `ICLR_New (2).pdf`   \n2. `542proposal.pdf` "
+              }
+            ]
+          }
+        ],
+        "forkedTraces": [],
+        "mergedTraces": [],
+        "response": "There are now **2 PDF files** available in this conversation:\n\n1. `ICLR_New (2).pdf`   \n2. `542proposal.pdf` ",
+        "status": "success",
+        "inputs": 1,
+        "activeTraceIds": [
+          "fork-node_1765320756261-1765413083937"
+        ],
+        "attachedFileIds": [
+          "2d43e6a5-2823-4564-8beb-fef2721056f3"
+        ],
+        "querySentAt": 1765415098761,
+        "responseReceivedAt": 1765415155872
+      },
+      "width": 200,
+      "height": 64,
+      "selected": true,
+      "positionAbsolute": {
+        "x": 407.983813671746,
+        "y": 459.30622770179616
+      },
+      "dragging": false
+    },
+    {
+      "id": "node_1765415177931",
+      "type": "llmNode",
+      "position": {
+        "x": 800,
+        "y": 100
+      },
+      "data": {
+        "label": "PDF Files",
+        "model": "gpt-5.1",
+        "temperature": 1,
+        "systemPrompt": "",
+        "userPrompt": "我来多做几轮测试。现在呢？",
+        "mergeStrategy": "smart",
+        "reasoningEffort": "medium",
+        "enableGoogleSearch": true,
+        "traces": [
+          {
+            "id": "fork-node_1765320756261-1765413083937",
+            "sourceNodeId": "node_1765320756261",
+            "color": "hsl(163, 70%, 60%)",
+            "messages": [
+              {
+                "id": "node_1765320756261-user",
+                "role": "user",
+                "content": "testing search_files. Please search for pdf files in this conversation and count the number."
+              },
+              {
+                "id": "node_1765320756261-assistant",
+                "role": "assistant",
+                "content": "I found **1 PDF file** in this conversation: `542proposal.pdf`. "
+              },
+              {
+                "id": "node_1765413048314-user",
+                "role": "user",
+                "content": "what about now?"
+              },
+              {
+                "id": "node_1765413048314-assistant",
+                "role": "assistant",
+                "content": "There are now **2 PDF files** available in this conversation:\n\n1. `ICLR_New (2).pdf`   \n2. `542proposal.pdf` "
+              }
+            ]
+          }
+        ],
+        "outgoingTraces": [
+          {
+            "id": "fork-node_1765320756261-1765413083937",
+            "sourceNodeId": "node_1765320756261",
+            "color": "hsl(163, 70%, 60%)",
+            "messages": [
+              {
+                "id": "node_1765320756261-user",
+                "role": "user",
+                "content": "testing search_files. Please search for pdf files in this conversation and count the number."
+              },
+              {
+                "id": "node_1765320756261-assistant",
+                "role": "assistant",
+                "content": "I found **1 PDF file** in this conversation: `542proposal.pdf`. "
+              },
+              {
+                "id": "node_1765413048314-user",
+                "role": "user",
+                "content": "what about now?"
+              },
+              {
+                "id": "node_1765413048314-assistant",
+                "role": "assistant",
+                "content": "There are now **2 PDF files** available in this conversation:\n\n1. `ICLR_New (2).pdf`   \n2. `542proposal.pdf` "
+              },
+              {
+                "id": "node_1765415177931-user",
+                "role": "user",
+                "content": "我来多做几轮测试。现在呢？"
+              },
+              {
+                "id": "node_1765415177931-assistant",
+                "role": "assistant",
+                "content": "现在一共存在 **2 个 PDF 文件**：\n\n1. `ICLR_New (2).pdf`   \n2. `542proposal.pdf` "
+              }
+            ]
+          }
+        ],
+        "forkedTraces": [],
+        "mergedTraces": [],
+        "activeTraceIds": [],
+        "attachedFileIds": [],
+        "response": "现在一共存在 **2 个 PDF 文件**：\n\n1. `ICLR_New (2).pdf`   \n2. `542proposal.pdf` ",
+        "status": "success",
+        "inputs": 1,
+        "querySentAt": 1765415177931,
+        "responseReceivedAt": 1765415177931
+      },
+      "width": 200,
+      "height": 64
+    },
+    {
+      "id": "node_1765415203978",
+      "type": "llmNode",
+      "position": {
+        "x": 1150,
+        "y": 100
+      },
+      "data": {
+        "label": "PDF Files",
+        "model": "gpt-5.1",
+        "temperature": 1,
+        "systemPrompt": "",
+        "userPrompt": "那么现在呢",
+        "mergeStrategy": "smart",
+        "reasoningEffort": "medium",
+        "enableGoogleSearch": true,
+        "traces": [
+          {
+            "id": "fork-node_1765320756261-1765413083937",
+            "sourceNodeId": "node_1765320756261",
+            "color": "hsl(163, 70%, 60%)",
+            "messages": [
+              {
+                "id": "node_1765320756261-user",
+                "role": "user",
+                "content": "testing search_files. Please search for pdf files in this conversation and count the number."
+              },
+              {
+                "id": "node_1765320756261-assistant",
+                "role": "assistant",
+                "content": "I found **1 PDF file** in this conversation: `542proposal.pdf`. "
+              },
+              {
+                "id": "node_1765413048314-user",
+                "role": "user",
+                "content": "what about now?"
+              },
+              {
+                "id": "node_1765413048314-assistant",
+                "role": "assistant",
+                "content": "There are now **2 PDF files** available in this conversation:\n\n1. `ICLR_New (2).pdf`   \n2. `542proposal.pdf` "
+              },
+              {
+                "id": "node_1765415177931-user",
+                "role": "user",
+                "content": "我来多做几轮测试。现在呢？"
+              },
+              {
+                "id": "node_1765415177931-assistant",
+                "role": "assistant",
+                "content": "现在一共存在 **2 个 PDF 文件**：\n\n1. `ICLR_New (2).pdf`   \n2. `542proposal.pdf` "
+              }
+            ]
+          }
+        ],
+        "outgoingTraces": [
+          {
+            "id": "fork-node_1765320756261-1765413083937",
+            "sourceNodeId": "node_1765320756261",
+            "color": "hsl(163, 70%, 60%)",
+            "messages": [
+              {
+                "id": "node_1765320756261-user",
+                "role": "user",
+                "content": "testing search_files. Please search for pdf files in this conversation and count the number."
+              },
+              {
+                "id": "node_1765320756261-assistant",
+                "role": "assistant",
+                "content": "I found **1 PDF file** in this conversation: `542proposal.pdf`. "
+              },
+              {
+                "id": "node_1765413048314-user",
+                "role": "user",
+                "content": "what about now?"
+              },
+              {
+                "id": "node_1765413048314-assistant",
+                "role": "assistant",
+                "content": "There are now **2 PDF files** available in this conversation:\n\n1. `ICLR_New (2).pdf`   \n2. `542proposal.pdf` "
+              },
+              {
+                "id": "node_1765415177931-user",
+                "role": "user",
+                "content": "我来多做几轮测试。现在呢？"
+              },
+              {
+                "id": "node_1765415177931-assistant",
+                "role": "assistant",
+                "content": "现在一共存在 **2 个 PDF 文件**：\n\n1. `ICLR_New (2).pdf`   \n2. `542proposal.pdf` "
+              },
+              {
+                "id": "node_1765415203978-user",
+                "role": "user",
+                "content": "那么现在呢"
+              },
+              {
+                "id": "node_1765415203978-assistant",
+                "role": "assistant",
+                "content": "现在一共有 **3 个 PDF 文件**：\n\n1. `survey_on_survey (3).pdf` fileciteturn0file0turn0file7  \n2. `ICLR_New (2).pdf` fileciteturn0file1turn0file8  \n3. `542proposal.pdf` "
+              }
+            ]
+          }
+        ],
+        "forkedTraces": [],
+        "mergedTraces": [],
+        "activeTraceIds": [],
+        "attachedFileIds": [],
+        "response": "现在一共有 **3 个 PDF 文件**：\n\n1. `survey_on_survey (3).pdf` fileciteturn0file0turn0file7  \n2. `ICLR_New (2).pdf` fileciteturn0file1turn0file8  \n3. `542proposal.pdf` ",
+        "status": "success",
+        "inputs": 1,
+        "querySentAt": 1765415203978,
+        "responseReceivedAt": 1765415203978
+      },
+      "width": 200,
+      "height": 64
+    },
+    {
+      "id": "node_1765415227098",
+      "type": "llmNode",
+      "position": {
+        "x": 1500,
+        "y": 100
+      },
+      "data": {
+        "label": "Citation Markers",
+        "model": "gpt-5.1",
+        "temperature": 1,
+        "systemPrompt": "",
+        "userPrompt": "”fileciteturn0file0turn0file7“这一串字符是什么？",
+        "mergeStrategy": "smart",
+        "reasoningEffort": "medium",
+        "enableGoogleSearch": true,
+        "traces": [
+          {
+            "id": "fork-node_1765320756261-1765413083937",
+            "sourceNodeId": "node_1765320756261",
+            "color": "hsl(163, 70%, 60%)",
+            "messages": [
+              {
+                "id": "node_1765320756261-user",
+                "role": "user",
+                "content": "testing search_files. Please search for pdf files in this conversation and count the number."
+              },
+              {
+                "id": "node_1765320756261-assistant",
+                "role": "assistant",
+                "content": "I found **1 PDF file** in this conversation: `542proposal.pdf`. "
+              },
+              {
+                "id": "node_1765413048314-user",
+                "role": "user",
+                "content": "what about now?"
+              },
+              {
+                "id": "node_1765413048314-assistant",
+                "role": "assistant",
+                "content": "There are now **2 PDF files** available in this conversation:\n\n1. `ICLR_New (2).pdf`   \n2. `542proposal.pdf` "
+              },
+              {
+                "id": "node_1765415177931-user",
+                "role": "user",
+                "content": "我来多做几轮测试。现在呢？"
+              },
+              {
+                "id": "node_1765415177931-assistant",
+                "role": "assistant",
+                "content": "现在一共存在 **2 个 PDF 文件**：\n\n1. `ICLR_New (2).pdf`   \n2. `542proposal.pdf` "
+              },
+              {
+                "id": "node_1765415203978-user",
+                "role": "user",
+                "content": "那么现在呢"
+              },
+              {
+                "id": "node_1765415203978-assistant",
+                "role": "assistant",
+                "content": "现在一共有 **3 个 PDF 文件**：\n\n1. `survey_on_survey (3).pdf` fileciteturn0file0turn0file7  \n2. `ICLR_New (2).pdf` fileciteturn0file1turn0file8  \n3. `542proposal.pdf` "
+              }
+            ]
+          }
+        ],
+        "outgoingTraces": [
+          {
+            "id": "fork-node_1765320756261-1765413083937",
+            "sourceNodeId": "node_1765320756261",
+            "color": "hsl(163, 70%, 60%)",
+            "messages": [
+              {
+                "id": "node_1765320756261-user",
+                "role": "user",
+                "content": "testing search_files. Please search for pdf files in this conversation and count the number."
+              },
+              {
+                "id": "node_1765320756261-assistant",
+                "role": "assistant",
+                "content": "I found **1 PDF file** in this conversation: `542proposal.pdf`. "
+              },
+              {
+                "id": "node_1765413048314-user",
+                "role": "user",
+                "content": "what about now?"
+              },
+              {
+                "id": "node_1765413048314-assistant",
+                "role": "assistant",
+                "content": "There are now **2 PDF files** available in this conversation:\n\n1. `ICLR_New (2).pdf`   \n2. `542proposal.pdf` "
+              },
+              {
+                "id": "node_1765415177931-user",
+                "role": "user",
+                "content": "我来多做几轮测试。现在呢？"
+              },
+              {
+                "id": "node_1765415177931-assistant",
+                "role": "assistant",
+                "content": "现在一共存在 **2 个 PDF 文件**：\n\n1. `ICLR_New (2).pdf`   \n2. `542proposal.pdf` "
+              },
+              {
+                "id": "node_1765415203978-user",
+                "role": "user",
+                "content": "那么现在呢"
+              },
+              {
+                "id": "node_1765415203978-assistant",
+                "role": "assistant",
+                "content": "现在一共有 **3 个 PDF 文件**：\n\n1. `survey_on_survey (3).pdf` fileciteturn0file0turn0file7  \n2. `ICLR_New (2).pdf` fileciteturn0file1turn0file8  \n3. `542proposal.pdf` "
+              },
+              {
+                "id": "node_1765415227098-user",
+                "role": "user",
+                "content": "”fileciteturn0file0turn0file7“这一串字符是什么？"
+              },
+              {
+                "id": "node_1765415227098-assistant",
+                "role": "assistant",
+                "content": "那一串  \n`fileciteturn0file0turn0file7`  \n\n是我用来**标注文件来源的内部引用标记（citation marker）**，含义大概是：\n\n- `filecite`：说明这是一个“文件引用”的开始  \n- `turn0file0`、`turn0file7`：指向本对话中由你上传的具体文件（第 0 轮的第 0 个文件、第 7 个文件之类的内部编号）  \n- ``：引用标记的结束符  \n\n这些标记主要是给系统用来“知道这句话是根据哪些文件内容回答的”，对你来说可以把它当成**文献引用**一样的东西，并不属于 PDF 文件的正文内容。"
+              }
+            ]
+          }
+        ],
+        "forkedTraces": [],
+        "mergedTraces": [],
+        "activeTraceIds": [],
+        "attachedFileIds": [],
+        "response": "那一串  \n`fileciteturn0file0turn0file7`  \n\n是我用来**标注文件来源的内部引用标记（citation marker）**，含义大概是：\n\n- `filecite`：说明这是一个“文件引用”的开始  \n- `turn0file0`、`turn0file7`：指向本对话中由你上传的具体文件（第 0 轮的第 0 个文件、第 7 个文件之类的内部编号）  \n- ``：引用标记的结束符  \n\n这些标记主要是给系统用来“知道这句话是根据哪些文件内容回答的”，对你来说可以把它当成**文献引用**一样的东西，并不属于 PDF 文件的正文内容。",
+        "status": "success",
+        "inputs": 1,
+        "querySentAt": 1765415227098,
+        "responseReceivedAt": 1765415227098
+      },
+      "width": 200,
+      "height": 64
+    },
+    {
+      "id": "node_1765415275416",
+      "type": "llmNode",
+      "position": {
+        "x": 1850,
+        "y": 100
+      },
+      "data": {
+        "label": "OpenAI Citation Rendering",
+        "model": "gpt-5.1",
+        "temperature": 1,
+        "systemPrompt": "",
+        "userPrompt": "openai平时是怎么渲染这个引用字符的？我在自己搭建前端",
+        "mergeStrategy": "smart",
+        "reasoningEffort": "medium",
+        "enableGoogleSearch": true,
+        "traces": [
+          {
+            "id": "fork-node_1765320756261-1765413083937",
+            "sourceNodeId": "node_1765320756261",
+            "color": "hsl(163, 70%, 60%)",
+            "messages": [
+              {
+                "id": "node_1765320756261-user",
+                "role": "user",
+                "content": "testing search_files. Please search for pdf files in this conversation and count the number."
+              },
+              {
+                "id": "node_1765320756261-assistant",
+                "role": "assistant",
+                "content": "I found **1 PDF file** in this conversation: `542proposal.pdf`. "
+              },
+              {
+                "id": "node_1765413048314-user",
+                "role": "user",
+                "content": "what about now?"
+              },
+              {
+                "id": "node_1765413048314-assistant",
+                "role": "assistant",
+                "content": "There are now **2 PDF files** available in this conversation:\n\n1. `ICLR_New (2).pdf`   \n2. `542proposal.pdf` "
+              },
+              {
+                "id": "node_1765415177931-user",
+                "role": "user",
+                "content": "我来多做几轮测试。现在呢？"
+              },
+              {
+                "id": "node_1765415177931-assistant",
+                "role": "assistant",
+                "content": "现在一共存在 **2 个 PDF 文件**：\n\n1. `ICLR_New (2).pdf`   \n2. `542proposal.pdf` "
+              },
+              {
+                "id": "node_1765415203978-user",
+                "role": "user",
+                "content": "那么现在呢"
+              },
+              {
+                "id": "node_1765415203978-assistant",
+                "role": "assistant",
+                "content": "现在一共有 **3 个 PDF 文件**：\n\n1. `survey_on_survey (3).pdf` fileciteturn0file0turn0file7  \n2. `ICLR_New (2).pdf` fileciteturn0file1turn0file8  \n3. `542proposal.pdf` "
+              },
+              {
+                "id": "node_1765415227098-user",
+                "role": "user",
+                "content": "”fileciteturn0file0turn0file7“这一串字符是什么？"
+              },
+              {
+                "id": "node_1765415227098-assistant",
+                "role": "assistant",
+                "content": "那一串  \n`fileciteturn0file0turn0file7`  \n\n是我用来**标注文件来源的内部引用标记（citation marker）**，含义大概是：\n\n- `filecite`：说明这是一个“文件引用”的开始  \n- `turn0file0`、`turn0file7`：指向本对话中由你上传的具体文件（第 0 轮的第 0 个文件、第 7 个文件之类的内部编号）  \n- ``：引用标记的结束符  \n\n这些标记主要是给系统用来“知道这句话是根据哪些文件内容回答的”，对你来说可以把它当成**文献引用**一样的东西，并不属于 PDF 文件的正文内容。"
+              }
+            ]
+          }
+        ],
+        "outgoingTraces": [],
+        "forkedTraces": [],
+        "mergedTraces": [],
+        "activeTraceIds": [],
+        "attachedFileIds": [],
+        "response": "简单说：  \n这些 `filecite...` 不应该直接展示给用户看，前端要**拦截并自己渲染成“引用小标签”**。\n\n虽然 OpenAI 没有公开一份“官方前端实现规范”，但在 ChatGPT 里，大致是这样处理的：\n\n1. **模型返回的原始文本里带有这串标记**：  \n   `…… survey_on_survey (3).pdf fileciteturn0file0turn0file7`  \n\n2. **前端在展示前做一次解析**：  \n   - 用正则把所有 `filecite ... ` 找出来  \n   - 拆成若干个 id：`turn0file0`, `turn0file7`  \n   - 根据这些 id 去你自己的文件元数据表里拿：文件名、类型、页码链接等  \n\n3. **把这段标记替换为你想要的 UI 元素**，比如：\n   - 一个小上标的引用号：`[1]`、`[2]`  \n   - 或一个带文件名的 tag / pill：`[survey_on_survey (3).pdf]`  \n   - 鼠标 hover / 点击后弹出：**文件名 + “打开文件 / 跳转到页码”按钮**\n\n示例实现思路（伪代码）：\n\n```ts\nconst CITATION_RE = /filecite([^]+)/g;\n\nfunction renderWithCitations(text, fileMetaMap) {\n  return text.replace(CITATION_RE, (match, inner) => {\n    const ids = inner.split(\"\"); // [\"turn0file0\", \"turn0file7\"]\n    // 这里你可以把 ids 映射成编号 [1] / [2]，或直接显示文件名\n    const label = \"[\" + ids.map(id => fileMetaMap[id]?.displayIndex ?? \"?\").join(\",\") + \"]\";\n    // 实际前端中可以渲染成组件，这里用纯文本代替\n    return label;\n  });\n}\n```\n\n所以在自己搭前端时，可以按这个思路做：\n\n- 把 `filecite...` 当成**纯结构化标记**  \n- 前端负责把它转换成：**引用编号 / tag / 弹窗等 UI**  \n- 用户最终不会看到这些奇怪字符，只看到你设计的引用样式即可。",
+        "status": "success",
+        "inputs": 1,
+        "querySentAt": 1765415275416,
+        "responseReceivedAt": 1765415275416
+      },
+      "width": 226,
+      "height": 64
+    }
+  ],
+  "edges": [
+    {
+      "source": "node_1765320756261",
+      "sourceHandle": "trace-fork-node_1765320756261-1765413083937",
+      "target": "node_1765413048314",
+      "targetHandle": "input-0",
+      "style": {
+        "stroke": "hsl(163, 70%, 60%)",
+        "strokeWidth": 2
+      },
+      "id": "reactflow__edge-node_1765320756261trace-fork-node_1765320756261-1765413083937-node_1765413048314input-0",
+      "data": {
+        "isMerged": false,
+        "colors": []
+      }
+    },
+    {
+      "source": "node_1765413048314",
+      "sourceHandle": "trace-fork-node_1765320756261-1765413083937",
+      "target": "node_1765415177931",
+      "targetHandle": "input-0",
+      "style": {
+        "stroke": "hsl(163, 70%, 60%)",
+        "strokeWidth": 2
+      },
+      "id": "reactflow__edge-node_1765413048314trace-fork-node_1765320756261-1765413083937-node_1765415177931input-0",
+      "data": {
+        "isMerged": false,
+        "colors": []
+      }
+    },
+    {
+      "source": "node_1765415177931",
+      "sourceHandle": "trace-fork-node_1765320756261-1765413083937",
+      "target": "node_1765415203978",
+      "targetHandle": "input-0",
+      "style": {
+        "stroke": "hsl(163, 70%, 60%)",
+        "strokeWidth": 2
+      },
+      "id": "reactflow__edge-node_1765415177931trace-fork-node_1765320756261-1765413083937-node_1765415203978input-0",
+      "data": {
+        "isMerged": false,
+        "colors": []
+      }
+    },
+    {
+      "source": "node_1765415203978",
+      "sourceHandle": "trace-fork-node_1765320756261-1765413083937",
+      "target": "node_1765415227098",
+      "targetHandle": "input-0",
+      "style": {
+        "stroke": "hsl(163, 70%, 60%)",
+        "strokeWidth": 2
+      },
+      "id": "reactflow__edge-node_1765415203978trace-fork-node_1765320756261-1765413083937-node_1765415227098input-0",
+      "data": {
+        "isMerged": false,
+        "colors": []
+      }
+    },
+    {
+      "source": "node_1765415227098",
+      "sourceHandle": "trace-fork-node_1765320756261-1765413083937",
+      "target": "node_1765415275416",
+      "targetHandle": "input-0",
+      "style": {
+        "stroke": "hsl(163, 70%, 60%)",
+        "strokeWidth": 2
+      },
+      "id": "reactflow__edge-node_1765415227098trace-fork-node_1765320756261-1765413083937-node_1765415275416input-0",
+      "data": {
+        "isMerged": false,
+        "colors": []
+      }
+    }
+  ],
+  "viewport": {
+    "x": 83.99242028051987,
+    "y": 146.6529258302924,
+    "zoom": 0.8705505632961249
+  },
+  "theme": "light"
+}
\ No newline at end of file