Upload public_routes.py

backend/media_analysis/public_routes.py

@@ -5,6 +5,7 @@ is per Python worker, not a distributed quota; add a gateway quota for productio
 """
 from __future__ import annotations
 import os
+import re
 import time
 from collections import defaultdict, deque
 from threading import Lock, BoundedSemaphore
@@ -18,6 +19,20 @@ _guard = Lock()
 _busy = BoundedSemaphore(value=1)
 
 
+
+
+def _public_answer_text(answer):
+    """Keep internal media/index keys in provenance, but not user-facing prose."""
+    if not isinstance(answer, str):
+        return answer
+    # A common model artifact: "Table 1 (paper 12010, label table_01)".
+    answer = re.sub(
+        r"\s*\(\s*paper\s*[:#]?\s*\d+\s*[,;]\s*label\s*[:#]?\s*(?:table|figure|fig)[_.-]0*\d+\s*\)",
+        "", answer, flags=re.I)
+    # Other formats of the same internal metadata are not useful to a reader.
+    answer = re.sub(r"\bpaper\s+\d+\s*[,;]\s*label\s+(?:table|figure|fig)[_.-]0*\d+\b", "", answer, flags=re.I)
+    return answer
+
 def _enabled():
     return os.getenv("WATERSCOPE_ENABLE_PUBLIC_MEDIA", "0").lower() in ("1", "true", "yes")
 
@@ -48,7 +63,6 @@ def media_analyze_public():
              "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.
-    import re
     if not re.search(r"\b(?:figure|fig\.?|table)\s*\d+\b", question, re.I):
         return jsonify({"error": "Specify Figure N or Table N"}), 400
     # Fixed provider and model on server, not client. Hard bound to avoid runaway bills.
@@ -76,12 +90,19 @@ def media_analyze_public():
                 except ValueError: pass
         return jsonify({"error": "Media analysis is busy; please retry shortly"}), 429
     try:
-        result = analyze_media_question(question.strip(), [paper], enabled=True,
+        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,
                  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")
         # Do not send large duplicate tabular records: CSV already carries the data.
         result.pop("structured_data", None)
+        if "answer" in result:
+            result["answer"] = _public_answer_text(result["answer"])
         result["export_csv_requested"] = export_csv
         return jsonify(result)
     except (MediaSelectionError, VisionConfigurationError, ValueError, FileNotFoundError) as exc: