Upload 2 files

metadata_search/agent_orchestrator.py

@@ -438,5 +438,5 @@ def agent_chat_payload(body: dict[str, Any]) -> dict[str, Any]:
     }
     if diagnostics is not None:
         payload["diagnostics"] = diagnostics
-    payload["llm_backend"] = llm_backend_name()
+    payload["llm_backend"] = "deepseek_fallback" if str(strategy).startswith("deepseek_fallback:") else llm_backend_name()
     return payload

metadata_search/chat_api.py

@@ -1,85 +1,106 @@
-"""Chat API for the WaterScope research agent (HF Space answers)."""
-
-from __future__ import annotations
-
-import os
-
-from backend.chat_settings import MAX_CHAT_PAPERS, enhance_ordinal_question, focus_document_ids_for_question
-from backend.mcp_sync import use_mcp
-from backend.model import llm_backend_name
-from backend.paper_voice import prefix_context, sanitize_answer
-from backend.waterscope_agent import agent_answer
-from metadata_search.mcp_retrieval import build_context_for_documents, list_papers_catalog, list_saved_papers
-
-
-def _generate_answer(
-    message: str,
-    context: str,
-    *,
-    application: str | None = None,
-    model: str | None = None,
-) -> tuple[str, str, str]:
-    answer, app_used, diagnostics = agent_answer(
-        prefix_context(context),
-        message,
-        application=application or os.getenv("WATERSCOPE_APPLICATION", "auto"),
-        model=model or os.getenv("WATERSCOPE_MODEL"),
-    )
-    return sanitize_answer(answer), f"waterscope:{app_used}", diagnostics
-
-
-def get_papers_payload() -> dict:
-    papers = list_papers_catalog()
-    saved_ids = {paper.get("document_id") for paper in list_saved_papers() if paper.get("document_id")}
-    for paper in papers:
-        paper["is_saved"] = paper.get("document_id") in saved_ids
-    return {"papers": papers, "max_select": MAX_CHAT_PAPERS}
-
-
-def chat_payload(body: dict) -> dict:
-    message = (body.get("message") or "").strip()
-    document_ids = body.get("document_ids") or []
-    application = body.get("application")
-    agent_mode = bool(body.get("agent_mode"))
-
-    if not message:
-        raise ValueError("message is required")
-
-    if agent_mode and not document_ids:
-        from metadata_search.agent_orchestrator import agent_chat_payload
-
-        return agent_chat_payload(body)
-
-    if not document_ids:
-        raise ValueError("Select at least one paper, or enable agent mode.")
-    if len(document_ids) > MAX_CHAT_PAPERS:
-        raise ValueError(f"Select at most {MAX_CHAT_PAPERS} papers.")
-
-    document_ids = focus_document_ids_for_question(message, document_ids)
-    context, sources = build_context_for_documents(document_ids, question=message)
-    message = enhance_ordinal_question(message, sources)
-
-    answer, strategy, diagnostics = _generate_answer(
-        message,
-        context,
-        application=application,
-        model=body.get("model"),
-    )
-
-    return {
-        "answer": answer,
-        "strategy": strategy,
-        "include_excerpts": False,
-        "sources": [
-            {
-                "document_id": source["document_id"],
-                "title": source["title"],
-                "file_name": source["file_name"],
-            }
-            for source in sources
-        ],
-        "chunks": [],
-        "retrieval": "mcp" if use_mcp() else "direct",
-        "diagnostics": diagnostics,
-        "llm_backend": llm_backend_name(),
-    }
+"""Chat API for the WaterScope research agent (HF Space answers)."""
+
+from __future__ import annotations
+
+import os
+
+from backend.chat_settings import MAX_CHAT_PAPERS, enhance_ordinal_question, focus_document_ids_for_question
+from backend.mcp_sync import use_mcp
+from backend.model import llm_backend_name
+from backend.paper_voice import prefix_context, sanitize_answer
+from backend.waterscope_agent import agent_answer
+from metadata_search.mcp_retrieval import build_context_for_documents, list_papers_catalog, list_saved_papers
+
+
+def _generate_answer(
+    message: str,
+    context: str,
+    *,
+    application: str | None = None,
+    model: str | None = None,
+) -> tuple[str, str, str]:
+    from backend.research_fallback import (
+        clear_llama_circuit, deepseek_research_answer, is_infrastructure_error,
+        llama_circuit_open, mark_llama_unavailable,
+    )
+    from backend.waterscope_agent import infer_application
+
+    requested_app = application or os.getenv("WATERSCOPE_APPLICATION", "auto")
+    app_used = infer_application(message, explicit=requested_app if requested_app != "auto" else None)
+    fallback_enabled = os.getenv("WATERSCOPE_ENABLE_RESEARCH_FALLBACK", "0") == "1"
+    if fallback_enabled and llama_circuit_open():
+        answer, fallback_model = deepseek_research_answer(prefix_context(context), message, app_used)
+        return sanitize_answer(answer), f"deepseek_fallback:{app_used}", f"provider=deepseek;model={fallback_model};reason=llama_circuit_open"
+
+    try:
+        answer, app_used, diagnostics = agent_answer(
+            prefix_context(context),
+            message,
+            application=requested_app,
+            model=model or os.getenv("WATERSCOPE_MODEL"),
+        )
+        clear_llama_circuit()
+        return sanitize_answer(answer), f"waterscope:{app_used}", diagnostics
+    except Exception as exc:
+        if not fallback_enabled or not is_infrastructure_error(exc):
+            raise
+        mark_llama_unavailable()
+        answer, fallback_model = deepseek_research_answer(prefix_context(context), message, app_used)
+        return sanitize_answer(answer), f"deepseek_fallback:{app_used}", f"provider=deepseek;model={fallback_model};reason=llama_unavailable"
+
+
+def get_papers_payload() -> dict:
+    papers = list_papers_catalog()
+    saved_ids = {paper.get("document_id") for paper in list_saved_papers() if paper.get("document_id")}
+    for paper in papers:
+        paper["is_saved"] = paper.get("document_id") in saved_ids
+    return {"papers": papers, "max_select": MAX_CHAT_PAPERS}
+
+
+def chat_payload(body: dict) -> dict:
+    message = (body.get("message") or "").strip()
+    document_ids = body.get("document_ids") or []
+    application = body.get("application")
+    agent_mode = bool(body.get("agent_mode"))
+
+    if not message:
+        raise ValueError("message is required")
+
+    if agent_mode and not document_ids:
+        from metadata_search.agent_orchestrator import agent_chat_payload
+
+        return agent_chat_payload(body)
+
+    if not document_ids:
+        raise ValueError("Select at least one paper, or enable agent mode.")
+    if len(document_ids) > MAX_CHAT_PAPERS:
+        raise ValueError(f"Select at most {MAX_CHAT_PAPERS} papers.")
+
+    document_ids = focus_document_ids_for_question(message, document_ids)
+    context, sources = build_context_for_documents(document_ids, question=message)
+    message = enhance_ordinal_question(message, sources)
+
+    answer, strategy, diagnostics = _generate_answer(
+        message,
+        context,
+        application=application,
+        model=body.get("model"),
+    )
+
+    return {
+        "answer": answer,
+        "strategy": strategy,
+        "include_excerpts": False,
+        "sources": [
+            {
+                "document_id": source["document_id"],
+                "title": source["title"],
+                "file_name": source["file_name"],
+            }
+            for source in sources
+        ],
+        "chunks": [],
+        "retrieval": "mcp" if use_mcp() else "direct",
+        "diagnostics": diagnostics,
+        "llm_backend": "deepseek_fallback" if strategy.startswith("deepseek_fallback:") else llm_backend_name(),
+    }