leocrm/app/plugins/builtins/unified_search/query_understanding.py

"""KI query understanding and result aggregation via LiteLLM."""

from __future__ import annotations

import json
import logging
import uuid
from typing import Any

from app.ai.llm_client import llm_complete
from sqlalchemy.ext.asyncio import AsyncSession

logger = logging.getLogger(__name__)

QUERY_ANALYZE_SYSTEM = (
    "Du bist ein Query-Analyzer fuer ein CRM. "
    "Analysiere die Suchanfrage und gib JSON zurueck: "
    '{"normalized_query": str, "entities": {"person": str|null, "contact": str|null, "topic": str|null}, '
    '"intent": str, "semantic_terms": [str], "suggested_filters": {}}'
)

RESULT_AGGREGATE_SYSTEM = (
    "Du bist ein Result-Aggregator. Fasse Ergebnisse zusammen und generiere Facetten: "
    '{"summary": str, "facets": {"types": {}, "dates": {}, "people": []}, "suggestions": [str]}'
)

# Sensible default model used when no provider-specific model is configured.
# llm_complete() resolves credentials/model prefix from the central config.
DEFAULT_LLM_MODEL = "ollama/deepseek-v4-flash"


def _fallback_query_analysis(query: str) -> dict[str, Any]:
    return {
        "normalized_query": query,
        "entities": {},
        "intent": "search",
        "semantic_terms": [],
        "suggested_filters": {},
    }


def _fallback_aggregate(results: list[dict], query: str) -> dict[str, Any]:
    return {
        "summary": f"{len(results)} Ergebnisse gefunden",
        "facets": {},
        "suggestions": [],
    }


def _parse_json_content(content: str) -> dict[str, Any]:
    """Parse LLM JSON response, stripping markdown code fences if present."""
    content = content.strip()
    if content.startswith("```"):
        content = content.split("\n", 1)[-1] if "\n" in content else content[3:]
    if content.endswith("```"):
        content = content[:-3].strip()
    return json.loads(content)


async def llm_analyze_query(
    query: str,
    db: AsyncSession | None = None,
    tenant_id: uuid.UUID | None = None,
) -> dict[str, Any]:
    """Analyze a search query using LLM for intent, entities, and semantic terms.

    Falls back to a simple dict if LLM fails.
    """
    try:
        result = await llm_complete(
            model=DEFAULT_LLM_MODEL,
            messages=[
                {"role": "system", "content": QUERY_ANALYZE_SYSTEM},
                {"role": "user", "content": query},
            ],
            temperature=0.1,
            max_tokens=500,
            response_format={"type": "json_object"},
            db=db,
            tenant_id=tenant_id,
        )
        return _parse_json_content(result["content"])
    except Exception:
        logger.warning("LLM query analysis failed, using fallback", exc_info=True)
        return _fallback_query_analysis(query)


async def llm_aggregate_results(
    results: list[dict],
    query: str,
    db: AsyncSession | None = None,
    tenant_id: uuid.UUID | None = None,
) -> dict[str, Any]:
    """Aggregate search results using LLM for summary, facets, and suggestions.

    Falls back to a simple dict if LLM fails.
    """
    if not results:
        return _fallback_aggregate(results, query)
    try:
        # Truncate results to avoid token overflow
        compact = [
            {"entity_type": r.get("entity_type"), "title": r.get("title", "")[:100]}
            for r in results[:50]
        ]
        user_msg = json.dumps({"query": query, "results": compact})

        result = await llm_complete(
            model=DEFAULT_LLM_MODEL,
            messages=[
                {"role": "system", "content": RESULT_AGGREGATE_SYSTEM},
                {"role": "user", "content": user_msg},
            ],
            temperature=0.1,
            max_tokens=1000,
            response_format={"type": "json_object"},
            db=db,
            tenant_id=tenant_id,
        )
        return _parse_json_content(result["content"])
    except Exception:
        logger.warning("LLM result aggregation failed, using fallback", exc_info=True)
        return _fallback_aggregate(results, query)