Update rag_retrieval.py

rag_retrieval.py

@@ -1,9 +1,9 @@
-# rag_retrieval.py
 import json
 import time
 import uuid
 import numpy as np
 import re
+from datetime import datetime
 
 from rag_config import RUNS_DIR, ROOT_DIR
 from rag_data import chunks, bm25, embeddings, sem_model, THREAD_OPTIONS
@@ -106,11 +106,10 @@ def retrieve_chunks(rewrite: str, session: dict, search_outside_thread: bool):
 
 def build_answer(user_text: str, rewrite: str, retrieved):
     """
-    Answer builder with a simple 'no clear answer' heuristic.
-
-    - If scores are very low OR none of the retrieved snippets share
-      meaningful words with the question, we return a graceful fallback.
-    - Otherwise, we list relevant snippets with citations.
+    Answer builder with:
+    - 'no clear answer' heuristic
+    - special handling for simple 'when' questions using email dates
+    - snippet list with citations for grounding
     """
     if not retrieved:
         return (
@@ -136,9 +135,42 @@ def build_answer(user_text: str, rewrite: str, retrieved):
             []
         )
 
-    # ---- Normal snippet-based answer ----
-    lines = [f"**Question:** {user_text}", "", "**Relevant information:**"]
+    # ---- Optional: direct answer for 'when' questions ----
+    direct_answer_line = None
+    if "when" in user_text.lower():
+        dated = []
+        for r in retrieved:
+            date_str = r.get("date")
+            if not date_str:
+                continue
+            try:
+                dt = datetime.fromisoformat(date_str.replace("Z", "+00:00"))
+                dated.append((dt, r))
+            except Exception:
+                continue
+
+        if dated:
+            # pick the latest email as the likely final approval/confirmation
+            dt_best, r_best = max(dated, key=lambda x: x[0])
+            nice_date = dt_best.strftime("%Y-%m-%d %H:%M")
+            direct_answer_line = (
+                f"**Answer:** The most relevant approval email in this thread "
+                f"was sent on **{nice_date}** "
+                f"[msg: {r_best['message_id']}]."
+            )
+
+    # ---- Build snippet-based explanation ----
+    lines = []
+    if direct_answer_line:
+        lines.append(direct_answer_line)
+        lines.append("")
+
+    lines.append(f"**Question:** {user_text}")
+    lines.append("")
+    lines.append("**Relevant information:**")
+
     citations = []
+    seen = set()  # avoid exact duplicate snippet+msg combos
 
     for r in retrieved:
         msg_id = r["message_id"]
@@ -146,6 +178,11 @@ def build_answer(user_text: str, rewrite: str, retrieved):
         snippet = r["text"].replace("\n", " ")
         snippet = (snippet[:300] + "…") if len(snippet) > 300 else snippet
 
+        key = (msg_id, snippet)
+        if key in seen:
+            continue
+        seen.add(key)
+
         if page_no is not None:
             cite = f"[msg: {msg_id}, page: {page_no}]"
         else: