ai.allucanget.biz/backend/app/services/models.py

"""Model cache service: fetch from OpenRouter, store in DuckDB."""
import json
from datetime import datetime, timedelta, timezone
from typing import Any

import duckdb

from . import openrouter

CACHE_TTL_HOURS = 24


def _normalize_modality(raw: str) -> str:
    """Normalize OpenRouter modality labels to canonical values."""
    value = (raw or "").strip().lower()
    if value in {"text", "image", "video", "audio", "embeddings", "embedding"}:
        return "embeddings" if value == "embedding" else value
    if "image" in value:
        return "image"
    if "video" in value:
        return "video"
    if "audio" in value:
        return "audio"
    if "embed" in value:
        return "embeddings"
    return "text"


def _parse_modality(raw_modality: str) -> str:
    """Extract output modality from OpenRouter architecture.modality string.

    Examples: "text->text", "text+image->text", "text->image", "text->video"
    """
    output = raw_modality.split(
        "->", 1)[-1] if "->" in raw_modality else raw_modality
    return _normalize_modality(output)


def _extract_output_modality(model: dict[str, Any]) -> str:
    """Extract output modality using OpenRouter schema, fallback to legacy field."""
    architecture = model.get("architecture") or {}

    output_modalities = architecture.get(
        "output_modalities") or model.get("output_modalities")
    if isinstance(output_modalities, list) and output_modalities:
        return _normalize_modality(str(output_modalities[0]))

    raw_modality = architecture.get(
        "modality") or model.get("modality") or "text->text"
    if isinstance(raw_modality, str):
        return _parse_modality(raw_modality)
    return "text"


async def _fetch_models_for_cache() -> list[dict[str, Any]]:
    """Fetch broad + modality-specific lists and merge unique models by id."""
    by_id: dict[str, dict[str, Any]] = {}

    # Primary fetch: all modalities (per OpenRouter docs).
    primary = await openrouter.list_models(output_modalities="all")
    for model in primary:
        model_id = model.get("id")
        if model_id:
            by_id[model_id] = model

    # Warmup fetches: some providers surface better results with explicit modality filter.
    for modality in ("image", "video", "audio", "embeddings", "text"):
        try:
            subset = await openrouter.list_models(output_modalities=modality)
        except Exception:
            continue
        for model in subset:
            model_id = model.get("id")
            if model_id and model_id not in by_id:
                by_id[model_id] = model

    return list(by_id.values())


async def refresh_models_cache(conn: duckdb.DuckDBPyConnection) -> int:
    """Fetch all models from OpenRouter and replace the cache. Returns count stored."""
    raw = await _fetch_models_for_cache()
    # Use naive UTC to avoid DuckDB TIMESTAMP tz-stripping inconsistencies
    now = datetime.now(timezone.utc).replace(tzinfo=None)

    conn.execute("DELETE FROM models_cache")
    count = 0
    for m in raw:
        modality = _extract_output_modality(m)
        pricing = m.get("pricing")
        model_id = m.get("id", "")
        if not model_id:
            continue
        conn.execute(
            """
            INSERT INTO models_cache (model_id, name, modality, context_length, pricing, fetched_at)
            VALUES (?, ?, ?, ?, ?, ?)
            ON CONFLICT (model_id) DO UPDATE SET
                name = excluded.name,
                modality = excluded.modality,
                context_length = excluded.context_length,
                pricing = excluded.pricing,
                fetched_at = excluded.fetched_at
            """,
            [
                model_id,
                m.get("name", model_id),
                modality,
                m.get("context_length"),
                json.dumps(pricing) if pricing else None,
                now,
            ],
        )
        count += 1
    return count


def is_cache_stale(conn: duckdb.DuckDBPyConnection) -> bool:
    """Return True if cache is empty or last fetched more than CACHE_TTL_HOURS ago."""
    row = conn.execute("SELECT MAX(fetched_at) FROM models_cache").fetchone()
    if not row or row[0] is None:
        return True
    last_fetched = row[0]
    # DuckDB TIMESTAMP is always naive; compare against naive UTC
    if last_fetched.tzinfo is not None:
        last_fetched = last_fetched.replace(tzinfo=None)
    now_naive = datetime.now(timezone.utc).replace(tzinfo=None)
    return now_naive - last_fetched > timedelta(hours=CACHE_TTL_HOURS)


def get_cached_models(
    conn: duckdb.DuckDBPyConnection,
    modality: str | None = None,
) -> list[dict[str, Any]]:
    """Return cached models, optionally filtered by modality, ordered by name."""
    if modality:
        rows = conn.execute(
            """
            SELECT model_id, name, modality, context_length, pricing
            FROM models_cache
            WHERE modality = ?
            ORDER BY name
            """,
            [modality],
        ).fetchall()
    else:
        rows = conn.execute(
            """
            SELECT model_id, name, modality, context_length, pricing
            FROM models_cache
            ORDER BY name
            """
        ).fetchall()

    result = []
    for row in rows:
        pricing = None
        if row[4]:
            try:
                pricing = json.loads(row[4])
            except (json.JSONDecodeError, TypeError):
                pricing = None
        result.append({
            "id": row[0],
            "name": row[1],
            "modality": row[2],
            "context_length": row[3],
            "pricing": pricing,
        })
    return result