Upload research_synthesis.py

metadata_search/research_synthesis.py

@@ -36,6 +36,7 @@ does NOT patch module globals, and therefore does not change production calls.
 from __future__ import annotations
 
 import math
+import hmac
 import os
 import time
 import types
@@ -283,7 +284,7 @@ def register_research_synthesis_route(research_bp):
         expected_token = os.getenv("WATERSCOPE_RESEARCH_TOKEN", "").strip()
         if expected_token:
             supplied = request.headers.get("X-Research-Token", "").strip()
-            if supplied != expected_token:
+            if not hmac.compare_digest(supplied, expected_token):
                 return jsonify({"ok": False, "error": "Unauthorized"}), 401
 
         body = request.get_json(force=True, silent=True) or {}
@@ -337,6 +338,29 @@ def register_research_synthesis_route(research_bp):
         application = _clean(body.get("application")) or "auto"
         model = _clean(body.get("model")) or None
 
+        # Optional research-only provider selection. Existing HF requests
+        # remain the default and keep their original auto-routing behavior.
+        provider = (_clean(body.get("provider")) or "hf").lower()
+        if provider not in {"hf", "azure"}:
+            return jsonify({"ok": False, "error": "provider must be hf or azure"}), 400
+        if provider == "azure":
+            if os.getenv("ENABLE_EXTERNAL_LLM", "false").lower() not in {"1", "true", "yes"}:
+                return jsonify({"ok": False, "error": "External LLM is disabled"}), 403
+            # A research-only secret is mandatory for paid Azure calls.
+            research_token = os.getenv("WATERSCOPE_RESEARCH_TOKEN", "").strip()
+            supplied_token = request.headers.get("X-Research-Token", "").strip()
+            if not research_token or not hmac.compare_digest(supplied_token, research_token):
+                return jsonify({"ok": False, "error": "Unauthorized"}), 401
+            # The HF 'auto' application detection lives in the model Space;
+            # external calls require explicit matching application prompts.
+            from llm_services.agent_prompts import AGENT_APPLICATIONS
+            if application not in AGENT_APPLICATIONS:
+                return jsonify({
+                    "ok": False,
+                    "error": "For Azure, specify an application: "
+                             + ", ".join(AGENT_APPLICATIONS),
+                }), 400
+
         try:
             max_words_per_paper = int(
                 body.get(
@@ -443,15 +467,40 @@ def register_research_synthesis_route(research_bp):
             )
 
             model_start = time.perf_counter()
-            answer, strategy, diagnostics = (
-                _research_generate_answer(
-                    question_to_model,
-                    full_context,
+            external_result = None
+            if provider == "azure":
+                from llm_services.llm_client import run_application
+                # The same complete context, question and application prompt
+                # are passed to both model paths. HF's prefix_context() is
+                # preserved in the external prompt as well.
+                external_result = run_application(
                     application=application,
-                    model=model,
-                    max_new_tokens=max_new_tokens_requested,
+                    context=model_context,
+                    question=question_to_model,
+                    provider="azure",
+                    model=None,  # use AZURE_LLM_DEPLOYMENT, not HF selector
+                    max_completion_tokens=max_new_tokens_requested,
+                )
+                answer = external_result["answer"]
+                strategy = f"waterscope:{application}"
+                diagnostics = {
+                    "provider": "azure",
+                    "model": external_result.get("model_returned"),
+                    "finish_reason": external_result.get("finish_reason"),
+                    "metering": external_result.get("metering"),
+                    "prompt_version": external_result.get("prompt_version"),
+                    "prompt_sha256": external_result.get("prompt_sha256"),
+                }
+            else:
+                answer, strategy, diagnostics = (
+                    _research_generate_answer(
+                        question_to_model,
+                        full_context,
+                        application=application,
+                        model=model,
+                        max_new_tokens=max_new_tokens_requested,
+                    )
                 )
-            )
             model_call_elapsed_s = (
                 time.perf_counter() - model_start
             )
@@ -474,6 +523,9 @@ def register_research_synthesis_route(research_bp):
                     answer_words > max_output_words
                 ),
                 "strategy": strategy,
+                "provider_used": provider,
+                "external_model": external_result.get("model_returned") if external_result else None,
+                "external_usage": external_result.get("metering") if external_result else None,
                 "application_requested": application,
                 "application_used": application_used,
                 "model_requested": model or "",
@@ -526,7 +578,7 @@ def register_research_synthesis_route(research_bp):
                     "mcp" if chat_api.use_mcp() else "direct"
                 ),
                 "diagnostics": diagnostics,
-                "llm_backend": chat_api.llm_backend_name(),
+                "llm_backend": "azure" if provider == "azure" else chat_api.llm_backend_name(),
                 "trace_status": {
                     "captured_full_context": bool(
                         full_context
@@ -555,7 +607,7 @@ def register_research_synthesis_route(research_bp):
                         MAX_RESEARCH_PAPERS
                     ),
                     "token_override_scope": (
-                        "isolated_function_clone"
+                        "external_api" if provider == "azure" else "isolated_function_clone"
                     ),
                 },
             }