server-26/drb-c2-core/app/internal/intelligence.py

"""
Gemini-powered intelligence extraction from call transcripts.

Sends the transcript to Gemini Flash with a tight JSON schema prompt.
Returns structured data: incident type, tags, location, vehicles, units, severity.

Falls back gracefully if Gemini is unavailable or returns malformed output.
"""
import asyncio
import json
from typing import Optional
from app.internal.logger import logger
from app.internal import firestore as fstore

_PROMPT_TEMPLATE = """You are analyzing a P25 public safety radio recording. The audio was transcribed by Whisper through a digital radio vocoder, which introduces errors. Each numbered transmission is a separate PTT press from a different radio. Extract structured information and respond ONLY with a single valid JSON object — no markdown, no explanation.

Schema:
{{
  "incident_type": one of "fire" | "ems" | "police" | "accident" | "other" | "unknown",
  "tags": [list of specific descriptive tags, max 6, e.g. "two-car mva", "property-damage-only", "working fire", "shots-fired"],
  "location": "most specific location string found, or empty string",
  "vehicles": [vehicle descriptions mentioned, e.g. "Hyundai Tucson", "black sedan"],
  "units": [unit IDs or officer numbers mentioned, e.g. "Unit 511", "Car 4"],
  "severity": one of "minor" | "moderate" | "major" | "unknown",
  "transcript_corrected": "corrected full transcript string, or null if no corrections needed"
}}

Rules:
- location: prefer intersections > addresses > mile markers > route+town > route alone > town alone. Empty string if none.
- tags: be specific and lowercase, hyphenated. Do not repeat incident_type as a tag.
- units: only identifiers explicitly mentioned, not inferred.
- Do not invent details not present in the transcript.
- transcript_corrected: fix only clear STT errors caused by vocoder distortion (e.g. "Several" → "10-4", misheard street names, garbled unit IDs). Use the back-and-forth context between transmissions to resolve ambiguities. Keep all radio language as-is — do NOT decode codes into plain English. Return null if the transcript looks accurate.

System: {system_id}
Talkgroup: {talkgroup_name}
{transcript_block}"""


async def extract_tags(
    call_id: str,
    transcript: str,
    talkgroup_name: Optional[str] = None,
    talkgroup_id: Optional[int] = None,
    system_id: Optional[str] = None,
    segments: Optional[list[dict]] = None,
) -> tuple[list[str], Optional[str], Optional[str]]:
    """
    Extract incident tags, type, location, and corrected transcript via Gemini.

    Returns:
        (tags, primary_type, location)

    Side-effect: updates calls/{call_id} in Firestore with tags, location,
    vehicles, units, severity, transcript_corrected; also stores the call embedding.
    """
    result = await asyncio.to_thread(_sync_extract, transcript, talkgroup_name, talkgroup_id, system_id, segments)

    tags: list[str] = result.get("tags") or []
    incident_type: Optional[str] = result.get("incident_type") or None
    location: Optional[str] = result.get("location") or None
    vehicles: list[str] = result.get("vehicles") or []
    units: list[str] = result.get("units") or []
    severity: str = result.get("severity") or "unknown"
    transcript_corrected: Optional[str] = result.get("transcript_corrected") or None

    if incident_type in ("unknown", "other", ""):
        incident_type = None

    # Store embedding alongside structured data
    embedding = await asyncio.to_thread(_sync_embed, _embed_text(transcript, incident_type))

    updates: dict = {
        "tags": tags,
        "severity": severity,
    }
    if location:
        updates["location"] = location
    if vehicles:
        updates["vehicles"] = vehicles
    if units:
        updates["units"] = units
    if embedding:
        updates["embedding"] = embedding
    if transcript_corrected:
        updates["transcript_corrected"] = transcript_corrected

    try:
        await fstore.doc_set("calls", call_id, updates)
    except Exception as e:
        logger.warning(f"Could not save intelligence for call {call_id}: {e}")

    logger.info(
        f"Intelligence: call {call_id} → type={incident_type}, "
        f"tags={tags}, location={location!r}, severity={severity}, "
        f"corrected={transcript_corrected is not None}"
    )
    return tags, incident_type, location


def _build_transcript_block(transcript: str, segments: Optional[list[dict]]) -> str:
    """Format transcript as numbered transmissions if segments are available."""
    if segments and len(segments) > 1:
        lines = [f"{i+1}. [{s['start']}s] {s['text']}" for i, s in enumerate(segments)]
        return f"Transmissions ({len(segments)}):\n" + "\n".join(lines)
    return f"Transcript:\n{transcript}"


def _sync_extract(
    transcript: str,
    talkgroup_name: Optional[str],
    talkgroup_id: Optional[int],
    system_id: Optional[str],
    segments: Optional[list[dict]],
) -> dict:
    """Call Gemini Flash and parse the JSON response."""
    from app.config import settings
    import google.generativeai as genai

    if not settings.gemini_api_key:
        logger.warning("GEMINI_API_KEY not set — intelligence extraction disabled.")
        return {}

    genai.configure(api_key=settings.gemini_api_key)
    model = genai.GenerativeModel(
        "gemini-2.5-flash-lite",
        generation_config={"response_mime_type": "application/json"},
    )

    tg = f"{talkgroup_name} (TGID {talkgroup_id})" if talkgroup_id else (talkgroup_name or "unknown")
    prompt = _PROMPT_TEMPLATE.format(
        transcript_block=_build_transcript_block(transcript, segments),
        talkgroup_name=tg,
        system_id=system_id or "unknown",
    )

    try:
        response = model.generate_content(prompt)
        return json.loads(response.text)
    except json.JSONDecodeError as e:
        logger.warning(f"Gemini returned non-JSON: {e}")
        return {}
    except Exception as e:
        logger.warning(f"Gemini extraction failed: {e}")
        return {}


def _sync_embed(text: str) -> Optional[list[float]]:
    """Generate a text-embedding-3-small vector for semantic similarity."""
    from app.config import settings
    from openai import OpenAI

    if not settings.openai_api_key:
        return None

    try:
        client = OpenAI(api_key=settings.openai_api_key)
        result = client.embeddings.create(
            model="text-embedding-3-small",
            input=text,
        )
        return result.data[0].embedding
    except Exception as e:
        logger.warning(f"Embedding generation failed: {e}")
        return None


def _embed_text(transcript: str, incident_type: Optional[str]) -> str:
    """Build the text string to embed — transcript + type context."""
    prefix = f"[{incident_type}] " if incident_type else ""
    return f"{prefix}{transcript}"