init

thirdeye/backend/agents/context_detector.py

@@ -0,0 +1,107 @@
+"""Context Detector Agent — auto-classifies group type from messages."""
+import logging
+from backend.providers import call_llm
+from backend.agents.json_utils import extract_json_object
+
+logger = logging.getLogger("thirdeye.agents.context_detector")
+
+SYSTEM_PROMPT = """You analyze a batch of messages from a Telegram group and determine what TYPE of group this is.
+
+CLASSIFY into exactly ONE:
+- "dev" — Software engineering team (code, PRs, deployments, bugs, tech stack)
+- "product" — Product/business team (features, users, metrics, roadmap, competitors)
+- "client" — Client/agency channel (deliverables, timelines, approvals, invoices)
+- "community" — Community/interest group (recommendations, events, local info, casual)
+
+Respond ONLY with valid JSON (no markdown, no backticks):
+{"detected_lens": "dev|product|client|community", "confidence": 0.0-1.0, "evidence": ["signal1", "signal2", "signal3"]}
+"""
+
+VALID_LENSES = {"dev", "product", "client", "community"}
+
+
+def _heuristic_detect_context(messages_text: str) -> dict:
+    """Rule-based fallback when LLM output is malformed/unavailable."""
+    text = (messages_text or "").lower()
+
+    lens_keywords = {
+        "dev": [
+            "bug", "deploy", "deployment", "api", "database", "schema", "postgres", "mongo",
+            "timeout", "endpoint", "pod", "pr", "code", "docker", "stack", "integration",
+        ],
+        "product": [
+            "feature", "roadmap", "user", "users", "client", "customers", "complain", "pain",
+            "prioritize", "priority", "enterprise", "competitor", "demo", "sso", "dark mode",
+            "mobile", "stability", "integration",
+        ],
+        "client": [
+            "invoice", "deadline", "deliverable", "approval", "sign-off", "scope", "payment",
+            "contract", "proposal", "timeline", "meeting",
+        ],
+        "community": [
+            "event", "meetup", "recommend", "anyone", "community", "local", "where can i",
+            "suggestion", "friends", "weekend",
+        ],
+    }
+
+    scores = {
+        lens: sum(text.count(keyword) for keyword in keywords)
+        for lens, keywords in lens_keywords.items()
+    }
+
+    best_lens = max(scores, key=scores.get)
+    best_score = scores[best_lens]
+    if best_score == 0:
+        best_lens = "dev"
+
+    evidence = [k for k in lens_keywords[best_lens] if k in text][:3]
+    confidence = min(0.95, 0.35 + 0.08 * best_score) if best_score > 0 else 0.0
+
+    return {
+        "detected_lens": best_lens,
+        "confidence": round(confidence, 2),
+        "evidence": evidence or ["heuristic_fallback"],
+    }
+
+
+async def detect_context(messages_text: str) -> dict:
+    """Detect group type from a batch of messages."""
+    messages = [
+        {"role": "system", "content": SYSTEM_PROMPT},
+        {"role": "user", "content": f"Classify this group based on these messages:\n\n{messages_text}"},
+    ]
+
+    try:
+        result = await call_llm(
+            "fast_large",
+            messages,
+            temperature=0.1,
+            max_tokens=300,
+            response_format={"type": "json_object"},
+        )
+        parsed = extract_json_object(result.get("content", ""))
+
+        detected_lens = str(parsed.get("detected_lens", "dev")).strip().lower()
+        if detected_lens not in VALID_LENSES:
+            detected_lens = "dev"
+
+        confidence = parsed.get("confidence", 0.5)
+        try:
+            confidence = float(confidence)
+        except (TypeError, ValueError):
+            confidence = 0.5
+
+        evidence = parsed.get("evidence", [])
+        if not isinstance(evidence, list):
+            evidence = [str(evidence)]
+
+        return {
+            "detected_lens": detected_lens,
+            "confidence": max(0.0, min(1.0, confidence)),
+            "evidence": [str(x) for x in evidence][:5],
+        }
+    except Exception as e:
+        logger.error(f"Context detection failed: {e}")
+        fallback = _heuristic_detect_context(messages_text)
+        fallback["evidence"] = fallback["evidence"] + ["detection_failed"]
+        return fallback