Upload 2 files

app.py

@@ -162,8 +162,11 @@ def system_instruction():
         "B. If the RETRIEVED CONTEXT does not contain enough information to answer, say so plainly: "
         "\"I couldn't find that in the reference documents provided for this tool. Please ask your care "
         "team.\" Do not guess or fill gaps from general knowledge.\n"
-        "C. When you use information from the context, attribute it naturally to the source document "
-        "name(s) shown in the context (e.g. \"According to the NCCN patient guideline ...\").\n"
+        "C. When you use information from the context, attribute it naturally to the document "
+        "TITLE(S) shown in each passage's [Document: ...] label — use that title, not a number, and "
+        "never say \"Source 1/2/3\" (e.g. \"According to the NCCN Guideline ...\" or \"The CAR T cell "
+        "Therapy — International Myeloma Foundation page notes ...\"). Use the title exactly as shown; "
+        "do not invent or embellish document names.\n"
         "D. Never invent drug names, doses, statistics, or study results that are not in the context.\n\n"
         "CONVERSATION RULES:\n"
         "1. When you ask a follow-up question, ask exactly ONE question per turn, 1-3 sentences, "

knowledge_base.py

@@ -177,6 +177,38 @@ def chunk_text(text, source, target_chars=TARGET_CHARS, overlap=OVERLAP):
     return chunks
 
 
+# ----------------------------------------------------------------------
+# Source title formatting (filename -> human-readable document title)
+# ----------------------------------------------------------------------
+def source_title(filename):
+    """Human-readable title from a source filename.
+
+    Only removes file extensions and export-timestamp suffixes and tidies
+    separators — it never invents words, so it cannot hallucinate a title. If
+    the cleaned result is a single token with no spaces (an accession/ID code
+    such as 's41467-022-31430-0' or 'nihms385546'), it is prefixed with
+    'Document ' so the model refers to a document rather than a bare code.
+    Always returns a non-empty string.
+    """
+    if not filename:
+        return "Reference document"
+    original = os.path.basename(str(filename))
+    name = re.sub(r"\.(pdf|png|jpe?g|txt|md)$", "", original, flags=re.IGNORECASE)
+    # strip trailing export/tracking timestamp suffix, e.g. "-07-13-2026_04_20_PM"
+    name = re.sub(r"[-_]\d{1,2}[-_]\d{1,2}[-_]\d{2,4}[_-]\d{1,2}[_-]\d{2}[_-][AP]M$", "", name)
+    # tidy separators
+    name = name.replace(" _ ", " — ")            # site/section separator used by several files
+    name = name.replace("+", " ").replace("_", " ")
+    # hyphens joining two letters -> spaces (leave digit-joining hyphens in IDs alone)
+    name = re.sub(r"(?<=[A-Za-z])-(?=[A-Za-z])", " ", name)
+    name = re.sub(r"\s{2,}", " ", name).strip(" -—")
+    if not name:
+        return original
+    if " " not in name:                          # single token => identifier/code
+        return f"Document {name}"
+    return name
+
+
 # ----------------------------------------------------------------------
 # Embedding (Gemini) with batching + disk cache
 # ----------------------------------------------------------------------
@@ -313,8 +345,10 @@ class KnowledgeBase:
                  "score": float(sims[i])} for i in idx]
 
     def context_block(self, retrieved):
-        """Format retrieved chunks as a numbered, source-attributed context block."""
+        """Format retrieved chunks as a source-attributed context block, labeling
+        each passage with the document's readable title (not 'Source N')."""
         parts = []
-        for i, r in enumerate(retrieved, 1):
-            parts.append(f"[Source {i}: {r['source']}]\n{r['text']}")
+        for r in retrieved:
+            title = source_title(r["source"])
+            parts.append(f"[Document: {title}]\n{r['text']}")
         return "\n\n---\n\n".join(parts)