From 347b391c6ebb49cc7b7cfd9ba4117f47b765d02a Mon Sep 17 00:00:00 2001
From: pyr0ball <pyroballpcs@gmail.com>
Date: Wed, 6 May 2026 10:17:51 -0700
Subject: [PATCH] fix: prevent LLM hallucination when retrieval returns
 low-signal results

- Strengthen synthesizer system prompt: hard 'respond with exactly' constraint
  instead of soft 'say so'; removes any wiggle room for the model to supplement
  from training data
- Add early return in synthesize() when chunks is empty (belt-and-suspenders
  alongside the existing guard in chat.py)
- Add MIN_SIGNAL threshold (0.01) in retriever: if the top combined score is
  below the threshold, return empty so the caller's no-results path fires instead
  of sending noise chunks to the LLM
---
 app/services/retriever.py   | 10 ++++++++++
 app/services/synthesizer.py | 19 +++++++++++++++----
 2 files changed, 25 insertions(+), 4 deletions(-)

diff --git a/app/services/retriever.py b/app/services/retriever.py
index 35ee597..20f0895 100644
--- a/app/services/retriever.py
+++ b/app/services/retriever.py
@@ -171,6 +171,13 @@ class Retriever:
             return bm25 * 0.5 + vec * 0.5
 
         ranked = sorted(merged.values(), key=_combined, reverse=True)[:top_k]
+        # Discard results where the best match is pure noise (neither BM25 term
+        # overlap nor vector similarity exceeded the minimum signal threshold).
+        # This lets the caller's empty-result guard fire instead of sending
+        # low-confidence chunks to the LLM where it fills gaps with training data.
+        MIN_SIGNAL = 0.01
+        if ranked and _combined(ranked[0]) < MIN_SIGNAL:
+            return []
         adjacent = _fetch_adjacent(ranked, db_path)
         return ranked + adjacent
 
@@ -189,5 +196,8 @@ class Retriever:
             )
             for r in self._bm25.query(query, top_k=top_k, doc_ids=doc_ids)
         ]
+        MIN_SIGNAL = 0.01
+        if hits and hits[0].bm25_score < MIN_SIGNAL:
+            return []
         adjacent = _fetch_adjacent(hits, db_path)
         return hits + adjacent
diff --git a/app/services/synthesizer.py b/app/services/synthesizer.py
index b9efdf2..befba3e 100644
--- a/app/services/synthesizer.py
+++ b/app/services/synthesizer.py
@@ -11,10 +11,18 @@ from dataclasses import dataclass
 from app.services.retriever import RetrievedChunk
 
 _SYSTEM_PROMPT = (
-    "You are a helpful document assistant. "
-    "Answer the user's question using ONLY the provided document excerpts. "
-    "For each claim, cite the source page as [p.N]. "
-    "If the excerpts are insufficient, say so. Do not invent information."
+    "You are a document assistant. "
+    "Answer questions using ONLY the document excerpts provided. "
+    "Cite every claim with the source page as [p.N]. "
+    "If the excerpts do not contain the answer, respond with exactly: "
+    "'I could not find an answer to that question in the indexed documents.' "
+    "Do NOT use knowledge from outside the provided excerpts. "
+    "Do NOT speculate, infer, or guess beyond what is explicitly stated."
+)
+
+_NO_RESULTS_ANSWER = (
+    "I could not find any relevant passages in the indexed documents for that question. "
+    "Try rephrasing, or check that the relevant document has been ingested."
 )
 
 
@@ -42,6 +50,9 @@ class Synthesizer:
         history: list[dict],
         chunks: list[RetrievedChunk],
     ) -> SynthesisResult:
+        if not chunks:
+            return SynthesisResult(answer=_NO_RESULTS_ANSWER, citations=())
+
         # 1500 chars (~300 words) per chunk: enough to capture definitions that
         # appear mid-paragraph without blowing past a 32k-context model's limit.
         context_parts = [f"[p.{c.page_number}]\n{c.text[:1500]}" for c in chunks]