Initial commit

5 files changed, 301 insertions, 0 deletions

diff --git a/kg_rag/rag_based_generation/Llama/run_drug_repurposing.py b/kg_rag/rag_based_generation/Llama/run_drug_repurposing.py
new file mode 100644
index 0000000..0b8d2f0
--- /dev/null
+++ b/kg_rag/rag_based_generation/Llama/run_drug_repurposing.py
@@ -0,0 +1,60 @@
+'''
+This script takes the drug repurposing style questions from the csv file and save the result as another csv file. 
+This script makes use of Llama model.
+Before running this script, make sure to configure the filepaths in config.yaml file.
+'''
+
+from langchain import PromptTemplate, LLMChain
+from kg_rag.utility import *
+import sys
+
+QUESTION_PATH = config_data["DRUG_REPURPOSING_PATH"]
+SYSTEM_PROMPT = system_prompts["DRUG_REPURPOSING"]
+CONTEXT_VOLUME = int(config_data["CONTEXT_VOLUME"])
+QUESTION_VS_CONTEXT_SIMILARITY_PERCENTILE_THRESHOLD = float(config_data["QUESTION_VS_CONTEXT_SIMILARITY_PERCENTILE_THRESHOLD"])
+QUESTION_VS_CONTEXT_MINIMUM_SIMILARITY = float(config_data["QUESTION_VS_CONTEXT_MINIMUM_SIMILARITY"])
+VECTOR_DB_PATH = config_data["VECTOR_DB_PATH"]
+NODE_CONTEXT_PATH = config_data["NODE_CONTEXT_PATH"]
+SENTENCE_EMBEDDING_MODEL_FOR_NODE_RETRIEVAL = config_data["SENTENCE_EMBEDDING_MODEL_FOR_NODE_RETRIEVAL"]
+SENTENCE_EMBEDDING_MODEL_FOR_CONTEXT_RETRIEVAL = config_data["SENTENCE_EMBEDDING_MODEL_FOR_CONTEXT_RETRIEVAL"]
+SAVE_PATH = config_data["SAVE_RESULTS_PATH"]
+MODEL_NAME = config_data["LLAMA_MODEL_NAME"]
+BRANCH_NAME = config_data["LLAMA_MODEL_BRANCH"]
+CACHE_DIR = config_data["LLM_CACHE_DIR"]
+
+
+save_name = "_".join(MODEL_NAME.split("/")[-1].split("-"))+"_drug_repurposing_questions_response.csv"
+
+
+INSTRUCTION = "Context:\n\n{context} \n\nQuestion: {question}"
+
+vectorstore = load_chroma(VECTOR_DB_PATH, SENTENCE_EMBEDDING_MODEL_FOR_NODE_RETRIEVAL)
+embedding_function_for_context_retrieval = load_sentence_transformer(SENTENCE_EMBEDDING_MODEL_FOR_CONTEXT_RETRIEVAL)
+node_context_df = pd.read_csv(NODE_CONTEXT_PATH)
+
+
+
+def main():
+    start_time = time.time()
+    llm = llama_model(MODEL_NAME, BRANCH_NAME, CACHE_DIR, max_new_tokens=1024)               
+    template = get_prompt(INSTRUCTION, SYSTEM_PROMPT)
+    prompt = PromptTemplate(template=template, input_variables=["context", "question"])
+    llm_chain = LLMChain(prompt=prompt, llm=llm)    
+    question_df = pd.read_csv(QUESTION_PATH)
+    answer_list = []
+    for index, row in question_df.iterrows():
+        question = row["text"]
+        context = retrieve_context(question, vectorstore, embedding_function_for_context_retrieval, node_context_df, CONTEXT_VOLUME, QUESTION_VS_CONTEXT_SIMILARITY_PERCENTILE_THRESHOLD, QUESTION_VS_CONTEXT_MINIMUM_SIMILARITY)
+        output = llm_chain.run(context=context, question=question)
+        answer_list.append((row["disease_in_question"], row["refDisease"], row["compoundGroundTruth"], row["text"], output))
+    answer_df = pd.DataFrame(answer_list, columns=["disease_in_question", "refDisease", "compoundGroundTruth", "text", "llm_answer"])
+    answer_df.to_csv(os.path.join(SAVE_PATH, save_name), index=False, header=True)
+    print("Completed in {} min".format((time.time()-start_time)/60))
+
+
+
+if __name__ == "__main__":
+    main()
+
+        
+
diff --git a/kg_rag/rag_based_generation/Llama/run_mcq_qa.py b/kg_rag/rag_based_generation/Llama/run_mcq_qa.py
new file mode 100644
index 0000000..67ae43c
--- /dev/null
+++ b/kg_rag/rag_based_generation/Llama/run_mcq_qa.py
@@ -0,0 +1,61 @@
+'''
+This script takes the MCQ style questions from the csv file and save the result as another csv file. 
+This script makes use of Llama model.
+Before running this script, make sure to configure the filepaths in config.yaml file.
+'''
+from tqdm import tqdm
+from langchain import PromptTemplate, LLMChain
+from kg_rag.utility import *
+
+
+QUESTION_PATH = config_data["MCQ_PATH"]
+SYSTEM_PROMPT = system_prompts["MCQ_QUESTION"]
+CONTEXT_VOLUME = int(config_data["CONTEXT_VOLUME"])
+QUESTION_VS_CONTEXT_SIMILARITY_PERCENTILE_THRESHOLD = float(config_data["QUESTION_VS_CONTEXT_SIMILARITY_PERCENTILE_THRESHOLD"])
+QUESTION_VS_CONTEXT_MINIMUM_SIMILARITY = float(config_data["QUESTION_VS_CONTEXT_MINIMUM_SIMILARITY"])
+VECTOR_DB_PATH = config_data["VECTOR_DB_PATH"]
+NODE_CONTEXT_PATH = config_data["NODE_CONTEXT_PATH"]
+SENTENCE_EMBEDDING_MODEL_FOR_NODE_RETRIEVAL = config_data["SENTENCE_EMBEDDING_MODEL_FOR_NODE_RETRIEVAL"]
+SENTENCE_EMBEDDING_MODEL_FOR_CONTEXT_RETRIEVAL = config_data["SENTENCE_EMBEDDING_MODEL_FOR_CONTEXT_RETRIEVAL"]
+SAVE_PATH = config_data["SAVE_RESULTS_PATH"]
+MODEL_NAME = config_data["LLAMA_MODEL_NAME"]
+BRANCH_NAME = config_data["LLAMA_MODEL_BRANCH"]
+CACHE_DIR = config_data["LLM_CACHE_DIR"]
+
+save_name = "_".join(MODEL_NAME.split("/")[-1].split("-"))+"_kg_rag_based_mcq_from_monarch_and_robokop_response.csv"
+
+
+INSTRUCTION = "Context:\n\n{context} \n\nQuestion: {question}"
+
+vectorstore = load_chroma(VECTOR_DB_PATH, SENTENCE_EMBEDDING_MODEL_FOR_NODE_RETRIEVAL)
+embedding_function_for_context_retrieval = load_sentence_transformer(SENTENCE_EMBEDDING_MODEL_FOR_CONTEXT_RETRIEVAL)
+node_context_df = pd.read_csv(NODE_CONTEXT_PATH)
+edge_evidence = False
+
+
+
+def main():    
+    start_time = time.time()
+    llm = llama_model(MODEL_NAME, BRANCH_NAME, CACHE_DIR)               
+    template = get_prompt(INSTRUCTION, SYSTEM_PROMPT)
+    prompt = PromptTemplate(template=template, input_variables=["context", "question"])
+    llm_chain = LLMChain(prompt=prompt, llm=llm)    
+    question_df = pd.read_csv(QUESTION_PATH)  
+    answer_list = []
+    for index, row in tqdm(question_df.iterrows()):
+        question = row["text"]
+        context = retrieve_context(question, vectorstore, embedding_function_for_context_retrieval, node_context_df, CONTEXT_VOLUME, QUESTION_VS_CONTEXT_SIMILARITY_PERCENTILE_THRESHOLD, QUESTION_VS_CONTEXT_MINIMUM_SIMILARITY, edge_evidence)
+        output = llm_chain.run(context=context, question=question)
+        answer_list.append((row["text"], row["correct_node"], output))
+    answer_df = pd.DataFrame(answer_list, columns=["question", "correct_answer", "llm_answer"])
+    answer_df.to_csv(os.path.join(SAVE_PATH, save_name), index=False, header=True) 
+    print("Completed in {} min".format((time.time()-start_time)/60))
+
+
+
+
+
+if __name__ == "__main__":
+    main()
+    
+    
diff --git a/kg_rag/rag_based_generation/Llama/run_mcq_qa_medgpt.py b/kg_rag/rag_based_generation/Llama/run_mcq_qa_medgpt.py
new file mode 100644
index 0000000..813b601
--- /dev/null
+++ b/kg_rag/rag_based_generation/Llama/run_mcq_qa_medgpt.py
@@ -0,0 +1,61 @@
+'''
+This script takes the MCQ style questions from the csv file and save the result as another csv file. 
+This script makes use of Llama model.
+Before running this script, make sure to configure the filepaths in config.yaml file.
+'''
+
+from langchain import PromptTemplate, LLMChain
+from kg_rag.utility import *
+
+
+QUESTION_PATH = config_data["MCQ_PATH"]
+SYSTEM_PROMPT = system_prompts["MCQ_QUESTION"]
+CONTEXT_VOLUME = int(config_data["CONTEXT_VOLUME"])
+QUESTION_VS_CONTEXT_SIMILARITY_PERCENTILE_THRESHOLD = float(config_data["QUESTION_VS_CONTEXT_SIMILARITY_PERCENTILE_THRESHOLD"])
+QUESTION_VS_CONTEXT_MINIMUM_SIMILARITY = float(config_data["QUESTION_VS_CONTEXT_MINIMUM_SIMILARITY"])
+VECTOR_DB_PATH = config_data["VECTOR_DB_PATH"]
+NODE_CONTEXT_PATH = config_data["NODE_CONTEXT_PATH"]
+SENTENCE_EMBEDDING_MODEL_FOR_NODE_RETRIEVAL = config_data["SENTENCE_EMBEDDING_MODEL_FOR_NODE_RETRIEVAL"]
+SENTENCE_EMBEDDING_MODEL_FOR_CONTEXT_RETRIEVAL = config_data["SENTENCE_EMBEDDING_MODEL_FOR_CONTEXT_RETRIEVAL"]
+SAVE_PATH = config_data["SAVE_RESULTS_PATH"]
+MODEL_NAME = 'PharMolix/BioMedGPT-LM-7B'
+BRANCH_NAME = 'main'
+CACHE_DIR = config_data["LLM_CACHE_DIR"]
+
+save_name = "_".join(MODEL_NAME.split("/")[-1].split("-"))+"_kg_rag_based_mcq_from_monarch_and_robokop_response.csv"
+
+
+INSTRUCTION = "Context:\n\n{context} \n\nQuestion: {question}"
+
+vectorstore = load_chroma(VECTOR_DB_PATH, SENTENCE_EMBEDDING_MODEL_FOR_NODE_RETRIEVAL)
+embedding_function_for_context_retrieval = load_sentence_transformer(SENTENCE_EMBEDDING_MODEL_FOR_CONTEXT_RETRIEVAL)
+node_context_df = pd.read_csv(NODE_CONTEXT_PATH)
+edge_evidence = False
+
+
+def main():    
+    start_time = time.time()
+    llm = llama_model(MODEL_NAME, BRANCH_NAME, CACHE_DIR)               
+    template = get_prompt(INSTRUCTION, SYSTEM_PROMPT)
+    prompt = PromptTemplate(template=template, input_variables=["context", "question"])
+    llm_chain = LLMChain(prompt=prompt, llm=llm)    
+    question_df = pd.read_csv(QUESTION_PATH)
+    question_df = question_df.sample(50, random_state=40)
+    answer_list = []
+    for index, row in question_df.iterrows():
+        question = row["text"]
+        context = retrieve_context(question, vectorstore, embedding_function_for_context_retrieval, node_context_df, CONTEXT_VOLUME, QUESTION_VS_CONTEXT_SIMILARITY_PERCENTILE_THRESHOLD, QUESTION_VS_CONTEXT_MINIMUM_SIMILARITY, edge_evidence)
+        output = llm_chain.run(context=context, question=question)
+        answer_list.append((row["text"], row["correct_node"], output))
+    answer_df = pd.DataFrame(answer_list, columns=["question", "correct_answer", "llm_answer"])
+    answer_df.to_csv(os.path.join(SAVE_PATH, save_name), index=False, header=True) 
+    print("Completed in {} min".format((time.time()-start_time)/60))
+
+
+
+
+
+if __name__ == "__main__":
+    main()
+    
+    
diff --git a/kg_rag/rag_based_generation/Llama/run_true_false_generation.py b/kg_rag/rag_based_generation/Llama/run_true_false_generation.py
new file mode 100644
index 0000000..fa1a37d
--- /dev/null
+++ b/kg_rag/rag_based_generation/Llama/run_true_false_generation.py
@@ -0,0 +1,59 @@
+'''
+This script takes the True/False style questions from the csv file and save the result as another csv file. 
+This script makes use of Llama model.
+Before running this script, make sure to configure the filepaths in config.yaml file.
+'''
+
+from langchain import PromptTemplate, LLMChain
+from kg_rag.utility import *
+import sys
+
+
+QUESTION_PATH = config_data["TRUE_FALSE_PATH"]
+SYSTEM_PROMPT = system_prompts["TRUE_FALSE_QUESTION"]
+QUESTION_VS_CONTEXT_SIMILARITY_PERCENTILE_THRESHOLD = float(config_data["QUESTION_VS_CONTEXT_SIMILARITY_PERCENTILE_THRESHOLD"])
+QUESTION_VS_CONTEXT_MINIMUM_SIMILARITY = float(config_data["QUESTION_VS_CONTEXT_MINIMUM_SIMILARITY"])
+VECTOR_DB_PATH = config_data["VECTOR_DB_PATH"]
+NODE_CONTEXT_PATH = config_data["NODE_CONTEXT_PATH"]
+SENTENCE_EMBEDDING_MODEL_FOR_NODE_RETRIEVAL = config_data["SENTENCE_EMBEDDING_MODEL_FOR_NODE_RETRIEVAL"]
+SENTENCE_EMBEDDING_MODEL_FOR_CONTEXT_RETRIEVAL = config_data["SENTENCE_EMBEDDING_MODEL_FOR_CONTEXT_RETRIEVAL"]
+SAVE_PATH = config_data["SAVE_RESULTS_PATH"]
+MODEL_NAME = config_data["LLAMA_MODEL_NAME"]
+BRANCH_NAME = config_data["LLAMA_MODEL_BRANCH"]
+CACHE_DIR = config_data["LLM_CACHE_DIR"]
+CONTEXT_VOLUME = 100
+edge_evidence = False
+
+save_name = "_".join(MODEL_NAME.split("/")[-1].split("-"))+"_kg_rag_based_true_false_binary_response.csv"
+
+
+INSTRUCTION = "Context:\n\n{context} \n\nQuestion: {question}"
+
+vectorstore = load_chroma(VECTOR_DB_PATH, SENTENCE_EMBEDDING_MODEL_FOR_NODE_RETRIEVAL)
+embedding_function_for_context_retrieval = load_sentence_transformer(SENTENCE_EMBEDDING_MODEL_FOR_CONTEXT_RETRIEVAL)
+node_context_df = pd.read_csv(NODE_CONTEXT_PATH)
+
+
+def main():    
+    start_time = time.time()
+    llm = llama_model(MODEL_NAME, BRANCH_NAME, CACHE_DIR)               
+    template = get_prompt(INSTRUCTION, SYSTEM_PROMPT)
+    prompt = PromptTemplate(template=template, input_variables=["context", "question"])
+    llm_chain = LLMChain(prompt=prompt, llm=llm)    
+    question_df = pd.read_csv(QUESTION_PATH)  
+    answer_list = []
+    for index, row in question_df.iterrows():
+        question = row["text"]
+        context = retrieve_context(question, vectorstore, embedding_function_for_context_retrieval, node_context_df, CONTEXT_VOLUME, QUESTION_VS_CONTEXT_SIMILARITY_PERCENTILE_THRESHOLD, QUESTION_VS_CONTEXT_MINIMUM_SIMILARITY, edge_evidence)
+        output = llm_chain.run(context=context, question=question)
+        answer_list.append((row["text"], row["label"], output))
+    answer_df = pd.DataFrame(answer_list, columns=["question", "label", "llm_answer"])
+    answer_df.to_csv(os.path.join(SAVE_PATH, save_name), index=False, header=True) 
+    print("Completed in {} min".format((time.time()-start_time)/60))
+
+    
+    
+if __name__ == "__main__":
+    main()
+    
+    
\ No newline at end of file
diff --git a/kg_rag/rag_based_generation/Llama/text_generation.py b/kg_rag/rag_based_generation/Llama/text_generation.py
new file mode 100644
index 0000000..2824135
--- /dev/null
+++ b/kg_rag/rag_based_generation/Llama/text_generation.py
@@ -0,0 +1,60 @@
+from langchain import PromptTemplate, LLMChain
+from kg_rag.utility import *
+import argparse
+
+
+
+parser = argparse.ArgumentParser()
+parser.add_argument('-i', type=bool, default=False, help='Flag for interactive mode')
+parser.add_argument('-m', type=str, default='method-1', help='Method to choose for Llama model')
+parser.add_argument('-e', type=bool, default=False, help='Flag for showing evidence of association from the graph')
+args = parser.parse_args()
+
+INTERACTIVE = args.i
+METHOD = args.m
+EDGE_EVIDENCE = bool(args.e)
+
+
+SYSTEM_PROMPT = system_prompts["KG_RAG_BASED_TEXT_GENERATION"]
+CONTEXT_VOLUME = int(config_data["CONTEXT_VOLUME"])
+QUESTION_VS_CONTEXT_SIMILARITY_PERCENTILE_THRESHOLD = float(config_data["QUESTION_VS_CONTEXT_SIMILARITY_PERCENTILE_THRESHOLD"])
+QUESTION_VS_CONTEXT_MINIMUM_SIMILARITY = float(config_data["QUESTION_VS_CONTEXT_MINIMUM_SIMILARITY"])
+VECTOR_DB_PATH = config_data["VECTOR_DB_PATH"]
+NODE_CONTEXT_PATH = config_data["NODE_CONTEXT_PATH"]
+SENTENCE_EMBEDDING_MODEL_FOR_NODE_RETRIEVAL = config_data["SENTENCE_EMBEDDING_MODEL_FOR_NODE_RETRIEVAL"]
+SENTENCE_EMBEDDING_MODEL_FOR_CONTEXT_RETRIEVAL = config_data["SENTENCE_EMBEDDING_MODEL_FOR_CONTEXT_RETRIEVAL"]
+MODEL_NAME = config_data["LLAMA_MODEL_NAME"]
+BRANCH_NAME = config_data["LLAMA_MODEL_BRANCH"]
+CACHE_DIR = config_data["LLM_CACHE_DIR"]
+
+
+INSTRUCTION = "Context:\n\n{context} \n\nQuestion: {question}"
+
+vectorstore = load_chroma(VECTOR_DB_PATH, SENTENCE_EMBEDDING_MODEL_FOR_NODE_RETRIEVAL)
+embedding_function_for_context_retrieval = load_sentence_transformer(SENTENCE_EMBEDDING_MODEL_FOR_CONTEXT_RETRIEVAL)
+node_context_df = pd.read_csv(NODE_CONTEXT_PATH)
+
+def main():
+    print(" ")
+    question = input("Enter your question : ")    
+    if not INTERACTIVE:
+        template = get_prompt(INSTRUCTION, SYSTEM_PROMPT)
+        prompt = PromptTemplate(template=template, input_variables=["context", "question"])
+        llm = llama_model(MODEL_NAME, BRANCH_NAME, CACHE_DIR, stream=True, method=METHOD) 
+        llm_chain = LLMChain(prompt=prompt, llm=llm)            
+        print("Retrieving context from SPOKE graph...")
+        context = retrieve_context(question, vectorstore, embedding_function_for_context_retrieval, node_context_df, CONTEXT_VOLUME, QUESTION_VS_CONTEXT_SIMILARITY_PERCENTILE_THRESHOLD, QUESTION_VS_CONTEXT_MINIMUM_SIMILARITY, EDGE_EVIDENCE)
+        print("Here is the KG-RAG based answer using Llama:")
+        print("")
+        output = llm_chain.run(context=context, question=question)
+    else:
+        interactive(question, vectorstore, node_context_df, embedding_function_for_context_retrieval, "llama", EDGE_EVIDENCE, SYSTEM_PROMPT, llama_method=METHOD)
+
+
+
+        
+
+
+
+if __name__ == "__main__":
+    main()