Initial release: VARS - personalized LLM with RAG and user vector learning

1 files changed, 25 insertions, 0 deletions

diff --git a/src/personalization/user_model/scoring.py b/src/personalization/user_model/scoring.py
new file mode 100644
index 0000000..75ffc84
--- /dev/null
+++ b/src/personalization/user_model/scoring.py
@@ -0,0 +1,25 @@
+import numpy as np
+from .tensor_store import UserState
+
+def score_with_user(
+    base_score: float,
+    user_state: UserState,
+    v_m: np.ndarray,        # [k]
+    beta_long: float,
+    beta_short: float,
+) -> float:
+    """
+    Personalized scoring:
+    s = base_score + (beta_long * z_long + beta_short * z_short) . v_m
+    Day2: beta_long = beta_short = 0 -> s == base_score
+    """
+    z_eff = beta_long * user_state.z_long + beta_short * user_state.z_short
+    # dot product
+    # Ensure shapes match
+    if v_m.shape != z_eff.shape:
+        # Just in case of dimension mismatch
+        return float(base_score)
+        
+    term = np.dot(z_eff, v_m)
+    return float(base_score + term)
+