Upload public_routes.py

backend/media_analysis/public_routes.py

@@ -62,9 +62,18 @@ def media_analyze_public():
     paper = {"paper_id": str(idx), "metadata_row_index": idx,
              "paper_title": str(row.get("Title") or row.get("title") or "")[:400],
              "doi": str(row.get("DOI") or row.get("doi") or "")[:200]}
-    # Match only explicit Figure N or Table N; media is never chosen arbitrarily.
-    if not re.search(r"\b(?:figure|fig\.?|table)\s*\d+\b", question, re.I):
+    # The resolver scans the entire question for numbered media labels. This
+    # public single-media endpoint must receive exactly ONE unique media label;
+    # instructional examples must never accidentally become requested sources.
+    refs = {
+        ("table" if m.group(1).lower() == "table" else "figure", int(m.group(2)))
+        for m in re.finditer(r"\b(fig(?:ure)?\.?|table)\s*0*(\d+)\b", question, re.I)
+    }
+    if not refs:
         return jsonify({"error": "Specify Figure N or Table N"}), 400
+    if len(refs) != 1:
+        return jsonify({"error": "Please analyze one numbered figure or table at a time"}), 400
+    requested_kind, requested_number = next(iter(refs))
     # Fixed provider and model on server, not client. Hard bound to avoid runaway bills.
     provider = os.getenv("WATERSCOPE_PUBLIC_MEDIA_PROVIDER", "deepseek").strip().lower()
     if provider not in ("azure", "deepseek"):
@@ -90,15 +99,23 @@ def media_analyze_public():
                 except ValueError: pass
         return jsonify({"error": "Media analysis is busy; please retry shortly"}), 429
     try:
-        model_question = question.strip() + (
-            "\n\nAnswer presentation: Use readable names (e.g., Table 1 or Figure 2). "
-            "Do not mention internal paper indexes, source labels like table_01, or "
-            "backend identifiers. Answer only what the question asks."
-        )
-        result = analyze_media_question(model_question, [paper], enabled=True,
+        # IMPORTANT: pass the user's media reference unchanged to the resolver.
+        # Appending example labels such as "Table 1 or Figure 2" makes the
+        # image selector load both sources, even for a single-table request.
+        result = analyze_media_question(question.strip(), [paper], enabled=True,
                  provider=provider, model=model, max_output_tokens=4096 if export_csv else 1600,
                  evidence_mode="combined", export_csv=export_csv,
                  analysis_mode="extract_all" if export_csv else "interpret")
+        # Fail closed if an upstream change ever introduces unrelated media.
+        sources = result.get("sources") or result.get("source_identifiers") or []
+        if len(sources) != 1:
+            return jsonify({"error": "Media source mismatch: expected exactly one source"}), 502
+        src_label = str(sources[0].get("label") or "")
+        src_match = re.search(r"\b(fig(?:ure)?|table)[\s_.-]*0*(\d+)\b", src_label, re.I)
+        if (not src_match or
+            ("table" if src_match.group(1).lower() == "table" else "figure") != requested_kind or
+            int(src_match.group(2)) != requested_number):
+            return jsonify({"error": "Media source mismatch: requested and returned labels differ"}), 502
         # Do not send large duplicate tabular records: CSV already carries the data.
         result.pop("structured_data", None)
         if "answer" in result: