Upload 4 files

backend/media_analysis/figure_analysis.py

@@ -117,7 +117,7 @@ def prepare_media_evidence(
 
 
 def analyze_media_question(question: str, selected_papers: list[dict], *, enabled: bool | None = None,
-                           figures_lookup=None, asset_resolver=None, vision_client=None) -> dict:
+                           figures_lookup=None, asset_resolver=None, vision_client=None, provider: str = "azure", model: str | None = None) -> dict:
     """Explicit opt-in entry point. Never silently substitutes a text summary answer."""
     if enabled is None:
         enabled = os.getenv("WATERSCOPE_ENABLE_MEDIA_ANALYSIS", "0").lower() in ("1", "true", "yes")
@@ -128,5 +128,5 @@ def analyze_media_question(question: str, selected_papers: list[dict], *, enable
     if vision_client is None:
         from .vision_client import analyze_images
         vision_client = analyze_images
-    result = vision_client(question, items)
+    result = vision_client(question, items, provider=provider, model=model)
     return {"status": "ok", "vision_inspected": True, **result}

backend/media_analysis/test_routes.py

@@ -33,6 +33,14 @@ def media_analyze_test():
     body = request.get_json(silent=True)
     if not isinstance(body, dict):
         return jsonify({"error": "JSON object required"}), 400
+    provider = str(body.get("provider", "azure")).strip().lower()
+    if provider not in ("azure", "deepseek"):
+        return jsonify({"error": "provider must be azure or deepseek"}), 400
+    model = body.get("model")
+    if model is not None:
+        if not isinstance(model, str) or not (1 <= len(model.strip()) <= 100):
+            return jsonify({"error": "model must be a nonempty string of at most 100 characters"}), 400
+        model = model.strip()
     question = body.get("question")
     papers = body.get("selected_papers")
     if not isinstance(question, str) or not question.strip() or len(question) > 4000:
@@ -51,7 +59,7 @@ def media_analyze_test():
             return jsonify({"error": "Paper summary exceeds limit"}), 400
         sanitized.append(clean)
     try:
-        result = analyze_media_question(question, sanitized, enabled=True)
+        result = analyze_media_question(question, sanitized, enabled=True, provider=provider, model=model)
         return jsonify(result)
     except (MediaSelectionError, VisionConfigurationError, ValueError, FileNotFoundError) as exc:
         return jsonify({"error": str(exc), "vision_inspected": False}), 400

backend/media_analysis/vision_client.py

@@ -1,4 +1,4 @@
-"""Opt-in Azure OpenAI Responses API client for scientific publication media.
+"""Opt-in Azure or DeepSeek Responses API client for scientific publication media.
 
 Dependencies: openai>=1.99 (a version with responses.create).
 Environment: existing AZURE_OPENAI_ENDPOINT, AZURE_LLM_DEPLOYMENT,
@@ -64,6 +64,7 @@ def _image_part(path: Path, detail: str = "high", max_bytes: int = 15_000_000):
 
 
 def analyze_images(question: str, evidence: list[dict], *, model: str | None = None,
+                   provider: str = "azure",
                    detail: str = "high", max_output_tokens: int = 1000,
                    client=None) -> dict:
     """Analyze pre-resolved evidence records containing image_path/caption/paper identifiers.
@@ -79,12 +80,25 @@ def analyze_images(question: str, evidence: list[dict], *, model: str | None = N
         raise ValueError("POC permits at most four images per request")
     if detail not in ("auto", "low", "high"):
         raise ValueError("detail must be auto, low or high")
-    model = model or os.getenv("AZURE_OPENAI_VISION_DEPLOYMENT") or _env_required("AZURE_LLM_DEPLOYMENT")
+    provider = str(provider).strip().lower()
+    if provider not in ("azure", "deepseek"):
+        raise ValueError("provider must be azure or deepseek")
+    if provider == "azure":
+        model = model or os.getenv("AZURE_OPENAI_VISION_DEPLOYMENT") or _env_required("AZURE_LLM_DEPLOYMENT")
+    else:
+        model = model or "deepseek-flash"
+        if model not in {"deepseek-flash", "deepseek-v4-pro"}:
+            raise ValueError("DeepSeek model must be deepseek-flash or deepseek-v4-pro")
     if client is None:
         from openai import OpenAI
-        client = OpenAI(base_url=_azure_responses_base_url(),
-                        api_key=_env_required("AZURE_OPENAI_API_KEY"),
-                        timeout=120.0, max_retries=1)
+        if provider == "azure":
+            client = OpenAI(base_url=_azure_responses_base_url(),
+                            api_key=_env_required("AZURE_OPENAI_API_KEY"),
+                            timeout=180.0, max_retries=1)
+        else:
+            client = OpenAI(base_url="https://api.deepseek.com",
+                            api_key=_env_required("DEEPSEEK_API_KEY"),
+                            timeout=180.0, max_retries=1)
 
     instructions = (
         "You analyze scientific publication figures and tables. Use only supplied images, "
@@ -122,7 +136,8 @@ def analyze_images(question: str, evidence: list[dict], *, model: str | None = N
     usage = getattr(response, "usage", None)
     return {
         "answer": getattr(response, "output_text", "") or "",
-        "strategy": "media_analysis:azure_responses",
+        "strategy": f"media_analysis:{provider}_responses",
+        "provider": provider,
         "model": model,
         "sources": used,
         "usage": {