ai insight rework #1

backend/apps/operations/services/insight_service.py

@@ -10,7 +10,11 @@ from __future__ import annotations
 import json
 import logging
 import re
-from typing import Any
+import threading
+import time
+import uuid
+from pathlib import Path
+from typing import Any, Callable
 
 import httpx
 from django.conf import settings
@@ -23,6 +27,9 @@ from apps.operations.services import root_cause
 
 logger = logging.getLogger(__name__)
 
+# Plan §5: one LLM inference at a time (CPU server; prevents thread thrash).
+_INFERENCE_LOCK = threading.Lock()
+
 _NUMERIC_PIPE_ROW = re.compile(r"^[\d.,]+(\s*\|\s*[\d.,]*)+$")
 
 # Stored in AIInsight.alert — condition of graded data, not narrative text.
@@ -434,7 +441,7 @@ def fetch_rag_chunks(query: str, topic: str, n_results: int = 4) -> tuple[list[s
         return [], []
 
 
-def call_ollama(system_prompt: str, user_prompt: str) -> str | None:
+def call_ollama(system_prompt: str, user_prompt: str) -> tuple[str | None, dict[str, Any]]:
     model = getattr(settings, "LLM_MODEL_NAME", "qwen2.5:3b") or "qwen2.5:3b"
     base = getattr(settings, "OLLAMA_BASE_URL", "http://127.0.0.1:11434") or "http://127.0.0.1:11434"
     url = f"{base.rstrip('/')}/api/chat"
@@ -442,7 +449,15 @@ def call_ollama(system_prompt: str, user_prompt: str) -> str | None:
     payload = {
         "model": model,
         "stream": False,
-        "options": {"temperature": 0.2},
+        # format:"json" (grammar) keeps the model from rambling in prose; the schema
+        # reminder appended AFTER the page-data JSON keeps it from copying that JSON.
+        "format": "json",
+        "options": {
+            "temperature": float(getattr(settings, "LLM_TEMPERATURE", 0.2) or 0.2),
+            "seed": int(getattr(settings, "LLM_SEED", 42) or 42),
+            "num_ctx": int(getattr(settings, "LLM_NUM_CTX", 2048) or 2048),
+            "num_thread": int(getattr(settings, "LLM_NUM_THREAD", 4) or 4),
+        },
         "messages": [
             {"role": "system", "content": system_prompt},
             {"role": "user", "content": user_prompt},
@@ -450,16 +465,23 @@ def call_ollama(system_prompt: str, user_prompt: str) -> str | None:
     }
     try:
         with httpx.Client(timeout=timeout) as client:
-            resp = client.post(url, json=payload)
+            # Plan §5: serialize LLM calls so concurrent generates don't thrash CPU.
+            with _INFERENCE_LOCK:
+                resp = client.post(url, json=payload)
             if resp.status_code >= 400:
                 logger.error("Ollama HTTP %s: %s", resp.status_code, resp.text[:500])
-                return None
+                return None, {}
             data = resp.json()
             message = data.get("message") or {}
-            return message.get("content") or data.get("response")
+            content = message.get("content") or data.get("response")
+            usage = {
+                "prompt_tokens": data.get("prompt_eval_count"),
+                "completion_tokens": data.get("eval_count"),
+            }
+            return content, usage
     except Exception as exc:  # noqa: BLE001
         logger.error("Ollama call failed: %s", exc)
-        return None
+        return None, {}
 
 
 def parse_llm_json(text: str) -> dict[str, str] | None:
@@ -530,23 +552,6 @@ def parse_end_cycle_llm_json(text: str) -> dict[str, Any] | None:
     return parsed
 
 
-def local_fallback_insight(graded: dict[str, Any], topic: str) -> dict[str, str]:
-    analyses = graded.get("analyses") or {}
-    lines = []
-    for name, block in analyses.items():
-        if isinstance(block, dict) and block.get("message"):
-            lines.append(f"{name}: {block['message']}")
-    if not lines:
-        return {
-            "kesimpulan": f"Data untuk topik {topic} tidak cukup untuk dianalisis (status unknown).",
-            "insight": "Lengkapi data operasional kandang, lalu generate ulang. Jangan mengarang angka.",
-        }
-    return {
-        "kesimpulan": lines[0],
-        "insight": "\n".join(lines[1:]) if len(lines) > 1 else lines[0],
-    }
-
-
 _SOFT_MORTALITY_WORD = re.compile(
     r"\b(rendah|normal|baik|aman|terkendali|sedikit)\b",
     re.IGNORECASE,
@@ -616,6 +621,11 @@ def collapse_duplicate_kandang(text: str) -> str:
     return _DUP_KANDANG.sub(lambda m: "Kandang" if m.group(0)[0].isupper() else "kandang", text)
 
 
+def _placeholder_text(value: str) -> bool:
+    """True when a narrative field is empty / schema placeholder (e.g. '...')."""
+    return len(value.strip().strip(".…<>- \t")) < 20
+
+
 def sanitize_insight_narrative(parsed: dict[str, Any]) -> dict[str, Any]:
     """Deterministic cleanup of narrative fields after the LLM."""
     out = dict(parsed)
@@ -736,6 +746,50 @@ def serialize_insight(row: AIInsight) -> dict[str, Any]:
     }
 
 
+# Plan Stage 1: graded status -> deterministic RAG keywords (no LLM query rewrite).
+_DIAGNOSTIC_KEYWORDS = {
+    "mortality": "penanganan mortalitas kumulatif deplesi penyebab kematian brooding",
+    "fcr": "target FCR pakan harian konsumsi pakan efisiensi ventilasi",
+    "bw": "target bobot badan ADG pertumbuhan standar mingguan",
+    "iot": "standar ventilasi suhu kelembapan amonia sekam basah kandang",
+    "environment": "standar ventilasi suhu kelembapan amonia sekam basah kandang",
+}
+
+
+def diagnose_rag_keywords(graded: dict[str, Any]) -> str:
+    """Stage 1: anomaly keywords from graded blocks; '' when everything ok/unknown."""
+    analyses = (graded or {}).get("analyses") or {}
+    clues: list[str] = []
+    for key, block in analyses.items():
+        if not isinstance(block, dict):
+            continue
+        if str(block.get("status") or "").lower() in ("warning", "critical"):
+            hint = _DIAGNOSTIC_KEYWORDS.get(key)
+            if hint and hint not in clues:
+                clues.append(hint)
+    return " ".join(clues)
+
+
+def _notify(stage_cb: Callable[[str], None] | None, stage: str) -> None:
+    if not stage_cb:
+        return
+    try:
+        stage_cb(stage)
+    except Exception:  # noqa: BLE001 - progress UI must never break generation
+        logger.debug("stage_cb(%s) failed", stage, exc_info=True)
+
+
+def _log_trace(record: dict[str, Any]) -> None:
+    """Plan Stage 5: append-only audit trail for replay/debug."""
+    try:
+        path = Path(settings.BASE_DIR) / "logs" / "rag_trace.jsonl"
+        path.parent.mkdir(parents=True, exist_ok=True)
+        with path.open("a", encoding="utf-8") as fh:
+            fh.write(json.dumps(record, ensure_ascii=False, default=str) + "\n")
+    except Exception:  # noqa: BLE001 - logging must never fail the request
+        logger.warning("rag_trace write failed", exc_info=True)
+
+
 def generate_insight(
     *,
     cycle_id: int,
@@ -745,6 +799,7 @@ def generate_insight(
     report_type: str = "page",
     report_period: str = "current",
     force_refresh: bool = False,
+    stage_cb: Callable[[str], None] | None = None,
 ) -> dict[str, Any]:
     if not kandang_id:
         raise ValueError("kandang_id wajib — insight tidak boleh untuk semua kandang")
@@ -785,7 +840,11 @@ def generate_insight(
     ctx.setdefault("cycleId", cycle_id)
     ctx = prune_context_by_period(ctx, report_type, report_period)
 
+    t_start = time.monotonic()
+    _notify(stage_cb, "grading")
     graded = grade_context(ctx)
+    t_after_grade = time.monotonic()
+    _notify(stage_cb, "retrieving")
     day = graded.get("hari_ke")
     standard_block = cp707.build_cp707_standard_block(ctx) or cp707.build_book_reference_context(day or 1)
 
@@ -805,7 +864,10 @@ def generate_insight(
         )
     else:
         rag_query = f"panduan manajemen broiler CP 707 untuk {topic} umur hari ke-{day or '?'}"
+    rag_query = f"{rag_query} {diagnose_rag_keywords(graded)}".strip()
     chunks, citations = fetch_rag_chunks(rag_query, topic)
+    t_after_rag = time.monotonic()
+    _notify(stage_cb, "synthesizing")
 
     system_prompt = (
         "Anda adalah asisten farm broiler on-premise. Tugas Anda HANYA menulis narasi "
@@ -829,8 +891,12 @@ def generate_insight(
             f"\n[NEXT CYCLE ACTIONS]\n"
             f"{json.dumps(next_actions, ensure_ascii=False, default=str)}\n"
         )
+    cited_chunks = [
+        f"[cp707#chunk{(citations[i].get('chunk_id') if i < len(citations) else None) or i}] {chunk}"
+        for i, chunk in enumerate(chunks[:4])
+    ]
     if chunks:
-        system_prompt += "\n[CUPLIKAN SOP CP 707 — prosa]\n" + "\n---\n".join(chunks[:4])
+        system_prompt += "\n[CUPLIKAN SOP CP 707 — prosa]\n" + "\n---\n".join(cited_chunks)
 
     user_prompt = build_insight_user_prompt(
         topic=topic,
@@ -839,44 +905,152 @@ def generate_insight(
         report_period=report_period,
         context=ctx,
     )
+    # Page-data JSON ends the user message; small models copy the last JSON they
+    # read. Re-state the required schema AFTER the data so recency wins.
+    # Status/facts go right before it: book chunks (system tail) describe IDEAL
+    # conditions, so without this last the model narrates ideals over actual status.
+    # Skip "unknown" blocks when at least one block has a real status:
+    # no-data lines are padding triggers (e.g. mortality spun into FCR
+    # narrative). All-unknown case keeps them: nothing factual to lean on
+    # and "data tidak tersedia" is then the only correct narration.
+    known_lines = [
+        f"{name} = {block.get('status')}: {block.get('message')}"
+        for name, block in (graded.get("analyses") or {}).items()
+        if isinstance(block, dict)
+        and block.get("message")
+        and block.get("status") != "unknown"
+    ]
+    unknown_lines = [
+        f"{name} = {block.get('status')}: {block.get('message')}"
+        for name, block in (graded.get("analyses") or {}).items()
+        if isinstance(block, dict)
+        and block.get("message")
+        and block.get("status") == "unknown"
+    ]
+    status_lines = known_lines or unknown_lines
+    actual_alert = alert_from_graded(graded)
+    user_prompt = (user_prompt or "").rstrip() + (
+        "\nKONTEKS: semua data adalah ternak AYAM BROILER (unggas) di kandang — "
+        "BUKAN tanaman/pertanian. 'Panen' = panen ayam. 'Bobot' = gram per ekor "
+        "(bukan per karung). 'ADG' = kenaikan bobot harian ayam."
+        f"\nSTATUS AKTUAL: {actual_alert}. Kesimpulan dan insight HARUS konsisten "
+        "dengan status ini: bila warning/critical, sebut masalahnya dan bandingkan "
+        "dengan angka standar CP 707; dilarang menyebut ideal/sehat/baik/aman "
+        "bila STATUS AKTUAL bukan healthy."
+        + ("\nFAKTA GRADED:\n- " + "\n- ".join(status_lines) if status_lines else "")
+        # No length req anywhere -> 3B model sometimes stops after 1 sentence
+        # (three EEF rows stored identical 136/127 chars). Demand structure.
+        + "\nPANJANG WAJIB: insight = narasi 3-6 kalimat berurutan: "
+        "(1) angka aktual vs standar CP 707, (2) penyebab/implikasi, "
+        "(3) tindakan konkret. kesimpulan = 1-3 kalimat. "
+        "Satu kalimat singkat = gagal."
+        # Placeholder <> (not "..."): model has copied schema "..." verbatim
+        # into insight, which stored fine and rendered as an empty page.
+        + '\nBALAS HANYA JSON persis: {"kesimpulan":"<ringkasan>","insight":"<narasi panjang>"}'
+        + ' - JANGAN salin teks <> dari contoh; tulis kalimat lengkap sendiri.'
+    )
 
-    raw = call_ollama(system_prompt, user_prompt)
-    if is_end_cycle:
-        parsed_raw = parse_end_cycle_llm_json(raw or "")
-        if parsed_raw:
-            # Apply mortality wording on narrative fields before sanitize.
-            soft = {
-                "kesimpulan": str(parsed_raw.get("kesimpulan") or ""),
-                "insight": str(parsed_raw.get("insight") or ""),
-            }
-            soft = enforce_mortality_wording(soft, graded)
-            soft = sanitize_insight_narrative(soft)
-            parsed_raw["kesimpulan"] = soft.get("kesimpulan")
-            parsed_raw["insight"] = soft.get("insight")
-            payload = root_cause.sanitize_end_cycle_payload(
-                parsed_raw,
-                hypotheses=hypotheses,
-                actions=next_actions,
-                masalah=masalah,
+    # Re-state output schema as the system prompt's final line (user_prompt
+    # repeats it again after the page-data JSON).
+    system_prompt += (
+        "\nINGAT: keluaran HANYA JSON valid sesuai FORMAT OUTPUT di atas "
+        "(wajib ada key kesimpulan dan insight). Tanpa teks lain."
+    )
+    # Degenerate output (schema placeholder like "..." copied verbatim) stored
+    # fine as JSON and rendered as an empty insight page. Reject + retry once,
+    # then fail loud (same policy as parse failure).
+    degenerate_reason: str | None = None
+    for attempt in range(2):
+        raw, usage = call_ollama(system_prompt, user_prompt)
+        t_after_llm = time.monotonic()
+        if is_end_cycle:
+            parsed_raw = parse_end_cycle_llm_json(raw or "")
+            if parsed_raw:
+                # Apply mortality wording on narrative fields before sanitize.
+                soft = {
+                    "kesimpulan": str(parsed_raw.get("kesimpulan") or ""),
+                    "insight": str(parsed_raw.get("insight") or ""),
+                }
+                soft = enforce_mortality_wording(soft, graded)
+                soft = sanitize_insight_narrative(soft)
+                parsed_raw["kesimpulan"] = soft.get("kesimpulan")
+                parsed_raw["insight"] = soft.get("insight")
+                payload = root_cause.sanitize_end_cycle_payload(
+                    parsed_raw,
+                    hypotheses=hypotheses,
+                    actions=next_actions,
+                    masalah=masalah,
+                )
+                payload = sanitize_insight_narrative(payload)
+            else:
+                # Fallback removed (2026-09-25): fail loudly instead of serving fake insight.
+                if not raw:
+                    raise RuntimeError("Ollama tidak mengembalikan output apa pun (cek layanan LLM).")
+                raise RuntimeError(
+                    "Output LLM end-cycle tidak sesuai format JSON yang diminta: "
+                    + repr(str(raw)[:200])
+                )
+            summary = collapse_duplicate_kandang(str(payload.get("kesimpulan") or ""))
+            insight_text = collapse_duplicate_kandang(
+                root_cause.format_end_cycle_insight_text(payload)
             )
-            payload = sanitize_insight_narrative(payload)
-        else:
-            payload = root_cause.local_fallback_end_cycle(graded, ctx)
-        summary = collapse_duplicate_kandang(str(payload.get("kesimpulan") or ""))
-        insight_text = collapse_duplicate_kandang(
-            root_cause.format_end_cycle_insight_text(payload)
-        )
-    else:
-        parsed = parse_llm_json(raw or "")
-        if not parsed:
-            parsed = local_fallback_insight(graded, topic)
         else:
+            parsed = parse_llm_json(raw or "")
+            if not parsed:
+                # Fallback removed (2026-09-25): fail loudly instead of serving fake insight.
+                if not raw:
+                    raise RuntimeError("Ollama tidak mengembalikan output apa pun (cek layanan LLM).")
+                raise RuntimeError(
+                    "Output LLM tidak sesuai format JSON yang diminta: " + repr(str(raw)[:200])
+                )
             parsed = enforce_mortality_wording(parsed, graded)
-        parsed = sanitize_insight_narrative(parsed)
-        insight_text = parsed["insight"]
-        summary = parsed["kesimpulan"]
+            parsed = sanitize_insight_narrative(parsed)
+            insight_text = parsed["insight"]
+            summary = parsed["kesimpulan"]
+        if _placeholder_text(summary) or _placeholder_text(insight_text):
+            degenerate_reason = f"kesimpulan={summary!r} insight={insight_text!r}"
+            user_prompt += (
+                "\nCATATAN: keluaran sebelumnya hanya placeholder. "
+                "Tulis kalimat lengkap sendiri untuk kesimpulan dan insight."
+            )
+            continue
+        degenerate_reason = None
+        break
+    if degenerate_reason:
+        raise RuntimeError(
+            "Output LLM degenerate (placeholder bukan kalimat): "
+            + degenerate_reason[:200]
+        )
 
     alert = alert_from_graded(graded)
+    _notify(stage_cb, "saving")
+    _log_trace(
+        {
+            "timestamp": timezone.now().isoformat(),
+            "request_id": str(uuid.uuid4()),
+            "cycle_id": cycle_id,
+            "kandang_name": kandang.kandang_name,
+            "topic": topic,
+            "report_type": report_type,
+            "report_period": report_period,
+            "model": getattr(settings, "LLM_MODEL_NAME", None),
+            "prompt_version": "v2.1",
+            "formulated_query": rag_query,
+            "retrieved_chunk_ids": [
+                f"cp707#{c.get('chunk_id')}" if c.get("chunk_id") is not None else f"cp707#{c.get('source')}"
+                for c in citations
+            ],
+            "graded_alert": alert,
+            "llm_ok": raw is not None,
+            "latency_breakdown": {
+                "grade_ms": round((t_after_grade - t_start) * 1000),
+                "retrieval_ms": round((t_after_rag - t_after_grade) * 1000),
+                "llm_ms": round((t_after_llm - t_after_rag) * 1000),
+                "total_ms": round((time.monotonic() - t_start) * 1000),
+            },
+            "token_usage": usage,
+        }
+    )
 
     row, _created = AIInsight.objects.update_or_create(
         cycle=cycle,