First Commit

src/paper_monitor/ai_enrich.py

@@ -0,0 +1,221 @@
+import json
+
+
+SYSTEM = '''You are a technical literature triage assistant for an automotive radar engineer.
+
+Analyze ONLY the supplied metadata.
+Do not invent experimental results, methods, conclusions, or claims that are not supported by the supplied metadata.
+
+Return:
+- relevance: integer from 1 to 5
+- categories: technical categories
+- summary: 1-2 concise Korean sentences
+- reason: concise Korean explanation of why this paper is relevant
+
+Preferred categories:
+Antenna,
+Waveform,
+DOA,
+MIMO,
+Target Modeling,
+Interference,
+RF Impairment,
+Imaging Radar,
+Signal Processing,
+Hardware,
+Other.
+'''
+
+
+AI_RESPONSE_SCHEMA = {
+    "type": "object",
+    "properties": {
+        "relevance": {
+            "type": "integer",
+            "minimum": 1,
+            "maximum": 5,
+        },
+        "categories": {
+            "type": "array",
+            "items": {
+                "type": "string"
+            },
+        },
+        "summary": {
+            "type": "string"
+        },
+        "reason": {
+            "type": "string"
+        },
+    },
+    "required": [
+        "relevance",
+        "categories",
+        "summary",
+        "reason",
+    ],    
+}
+
+
+def _make_payload(paper):
+    return {
+        "title": paper.title,
+        "abstract": paper.abstract,
+        "venue": paper.venue,
+        "year": paper.year,
+        "current_categories": paper.categories,
+        "current_relevance": paper.relevance,
+    }
+
+
+def _apply_result(paper, data):
+    paper.relevance = max(
+        1,
+        min(
+            5,
+            int(data.get("relevance", paper.relevance))
+        )
+    )
+
+    if isinstance(data.get("categories"), list) and data["categories"]:
+        paper.categories = [
+            str(x)
+            for x in data["categories"][:5]
+        ]
+
+    paper.summary = str(
+        data.get("summary", "")
+    ).strip()
+
+    paper.ai_reason = str(
+        data.get("reason", "")
+    ).strip()
+
+    return paper
+
+
+def enrich_with_gemini(paper, api_key, model):
+    if not api_key:
+        return paper
+
+    from google import genai
+    from google.genai import types
+
+    client = genai.Client(api_key=api_key)
+
+    payload = _make_payload(paper)
+
+    prompt = (
+        SYSTEM
+        + "\n\nPaper metadata:\n"
+        + json.dumps(payload, ensure_ascii=False)
+    )
+
+    response = client.models.generate_content(
+        model=model,
+        contents=prompt,
+        config=types.GenerateContentConfig(
+            temperature=0.1,
+            response_mime_type="application/json",
+            response_schema=AI_RESPONSE_SCHEMA,
+        ),
+    )
+
+    data = json.loads(response.text)
+
+    paper = _apply_result(paper, data)
+    paper.ai_analysis_level = "abstract"
+
+    return paper
+
+def enrich_with_gemini_full_text(
+    paper,
+    full_text,
+    api_key,
+    model,
+):
+    if not api_key:
+        return paper
+
+    if not full_text or not full_text.strip():
+        return paper
+
+    from google import genai
+    from google.genai import types
+
+    client = genai.Client(api_key=api_key)
+
+    payload = {
+        "title": paper.title,
+        "abstract": paper.abstract,
+        "venue": paper.venue,
+        "year": paper.year,
+        "full_text": full_text,
+        "current_categories": paper.categories,
+        "current_relevance": paper.relevance,
+    }
+
+    prompt = (
+        SYSTEM
+        + """
+
+You are given extracted full text from the paper.
+
+Base the analysis primarily on the supplied full text.
+Do not invent information that is not present in the supplied text.
+If the extracted text appears incomplete or corrupted, be conservative.
+
+The summary should describe the paper's main technical approach,
+contribution, and relevance to automotive radar when supported by the text.
+"""
+        + "\n\nPaper data:\n"
+        + json.dumps(payload, ensure_ascii=False)
+    )
+
+    response = client.models.generate_content(
+        model=model,
+        contents=prompt,
+        config=types.GenerateContentConfig(
+            temperature=0.1,
+            response_mime_type="application/json",
+            response_schema=AI_RESPONSE_SCHEMA,
+        ),
+    )
+
+    data = json.loads(response.text)
+
+    paper = _apply_result(paper, data)
+    paper.ai_analysis_level = "full_text"
+
+    return paper
+
+def enrich_with_openai(paper, api_key, model):
+    if not api_key:
+        return paper
+
+    from openai import OpenAI
+
+    client = OpenAI(api_key=api_key)
+
+    payload = _make_payload(paper)
+
+    response = client.responses.create(
+        model=model,
+        instructions=SYSTEM,
+        input=json.dumps(
+            payload,
+            ensure_ascii=False
+        ),
+    )
+
+    text = response.output_text.strip()
+
+    if text.startswith("```"):
+        text = text.strip("`")
+
+        if text.lower().startswith("json"):
+            text = text[4:].strip()
+
+    data = json.loads(text)
+
+    return _apply_result(paper, data)