Skip to content

Framework for orchestrating role-playing, autonomous AI agents. By fostering collaborative intelligence, CrewAI empowers agents to work together seamlessly, tackling complex tasks.

24.3 KB
Raw Blame
"""Dynamic model catalog for the crew-creation wizard.

Resolves the models to offer for a given provider using a three-tier strategy:

1. **Vendor API** - when the provider's API key is already present in the
   environment, query the vendor's own model-listing endpoint. This is the only
   source that reliably reflects the *latest* models (real release dates /
   display names, straight from the vendor).
2. **Curated hardcoded fallback** - the hand-verified list baked into the
   wizard, used when no API key is available. Authoritative but frozen, so it is
   refreshed periodically.
3. **LiteLLM feed** - the community ``model_prices_and_context_window.json`` the
   CLI already caches. Only used for providers with *no* curated list: the feed
   lags real releases badly (it can miss a vendor's newest models entirely), so
   it must never preempt the curated fallback.

Every tier is best-effort: any network error, timeout, missing key, or empty
result quietly falls through to the next tier, and the caller's hardcoded list
is always the final backstop. The picker never blocks for long — network calls
use a short timeout and successful results are cached.
"""

from __future__ import annotations

from collections.abc import Callable
import contextlib
import hashlib
import json
import os
from pathlib import Path
import re
import time
from typing import Any

import certifi
import httpx

from crewai_cli.constants import JSON_URL


# ── Tunables ─────────────────────────────────────────────────────

#: How many models to surface per provider.
MAX_MODELS = 8

#: Timeout (seconds) for any network call made while resolving models.
_TIMEOUT = 6.0

#: How long a resolved (dynamic) catalog stays fresh before we refetch. Kept
#: short: it only spares the picker repeated fetches within a wizard session,
#: and a stale list (new/removed models, account changes) is worse than a ~1s
#: refetch. Local providers (Ollama) are not cached at all — see _is_cacheable.
_CATALOG_TTL = 300

#: How long a fallback result is cached after a failed/empty fetch. Short, so a
#: newly-added API key takes effect soon, but long enough to spare the picker a
#: repeated timeout-prone network attempt on every call within one session.
_NEGATIVE_TTL = 300

#: How long the shared LiteLLM feed cache stays fresh.
_LITELLM_TTL = 24 * 3600

#: Env vars that may hold each provider's API key, in priority order. A
#: provider with an empty tuple (e.g. local Ollama) needs no key. Gemini accepts
#: either name, matching crewai's own Gemini provider.
_PROVIDER_KEY_ENV: dict[str, tuple[str, ...]] = {
    "openai": ("OPENAI_API_KEY",),
    "anthropic": ("ANTHROPIC_API_KEY",),
    "gemini": ("GEMINI_API_KEY", "GOOGLE_API_KEY"),
    "groq": ("GROQ_API_KEY",),
    "cerebras": ("CEREBRAS_API_KEY",),
    "ollama": (),
}


def _provider_api_key(provider_key: str) -> str | None:
    """First non-empty API key found among the provider's env vars."""
    for env in _PROVIDER_KEY_ENV.get(provider_key, ()):
        value = os.environ.get(env)
        if value:
            return value
    return None


# Substrings that mark a model id as *not* a chat/completion model. Used to
# filter noisy OpenAI-compatible ``/models`` listings.
_NON_CHAT_MARKERS = (
    "embedding",
    "embed",
    "whisper",
    "tts",
    "audio",
    "transcribe",
    "realtime",
    "dall-e",
    "dalle",
    "image",
    "moderation",
    "similarity",
    "-edit",
    "davinci-002",
    "babbage-002",
    "computer-use",
    "guard",
)

_ACRONYMS = {
    "gpt": "GPT",
    "ai": "AI",
    "nim": "NIM",
    "llm": "LLM",
    "hd": "HD",
    "us": "US",
    "eu": "EU",
    "oss": "OSS",
    "it": "IT",
}

# Tokens with non-title-case brand capitalization.
_BRAND_TOKENS = {
    "deepseek": "DeepSeek",
    "chatgpt": "ChatGPT",
    "qwq": "QwQ",
}


# ── Public API ───────────────────────────────────────────────────


def get_provider_models(
    provider_key: str, fallback: list[tuple[str, str]]
) -> list[tuple[str, str]]:
    """Return ``(model_id, label)`` pairs for ``provider_key``, newest first.

    Tries the vendor API (if a key is in the environment) first, since it is the
    only reliably-fresh source. When no key is available it returns the curated
    ``fallback`` verbatim — the LiteLLM feed is consulted **only** for providers
    with no curated list, because the feed lags real releases and would
    otherwise surface a staler list than the hand-verified fallback. Never
    raises: any failure degrades to the next tier.

    Args:
        provider_key: Short provider identifier, e.g. ``"anthropic"``.
        fallback: Curated ``(model_id, label)`` pairs to use as the backstop and
            to source friendly labels for known models.

    Returns:
        Up to :data:`MAX_MODELS` ``(model_id, label)`` pairs. Falls back to
        ``fallback`` verbatim when no fresher list can be resolved.
    """
    cached = _read_catalog_cache(provider_key)
    if cached is not None:
        return cached

    label_map = {model_id: label for model_id, label in fallback}

    # A non-None vendor result is authoritative — even when empty (e.g. a
    # reachable Ollama with no models installed): show that rather than
    # hardcoded suggestions the crew can't actually run. The picker handles an
    # empty list by prompting for manual entry.
    vendor = _from_vendor(provider_key)
    if vendor is not None:
        result = _finalize(vendor, label_map)
        if result:
            _write_catalog_cache(provider_key, result, source="dynamic")
        return result

    # Vendor tier unavailable. The LiteLLM feed lags real releases, so only
    # reach for it when we have no curated fallback — never override the fallback.
    entries = _from_litellm(provider_key) if not fallback else None
    result = _finalize(entries, label_map) if entries else []
    if result:
        _write_catalog_cache(provider_key, result, source="dynamic")
        return result

    # Nothing fresher than the curated list. Cache it briefly (negative cache)
    # so a failed vendor/LiteLLM fetch isn't retried on every subsequent call.
    # (_write_catalog_cache skips non-cacheable providers like Ollama.)
    if fallback:
        _write_catalog_cache(provider_key, fallback, source="fallback")
    return fallback


# ── Tier 1: vendor APIs ──────────────────────────────────────────


def _from_vendor(provider_key: str) -> list[dict[str, Any]] | None:
    """Fetch models from the vendor.

    Returns the model list on a successful fetch — **including an empty list**,
    which is meaningful (e.g. a reachable Ollama server with nothing installed).
    Returns ``None`` only when the vendor tier is unavailable: no fetcher, no
    API key, or the request failed.
    """
    fetcher = _VENDOR_FETCHERS.get(provider_key)
    if fetcher is None:
        return None

    api_key = _provider_api_key(provider_key)
    if _PROVIDER_KEY_ENV.get(provider_key) and not api_key:
        # Provider needs a key and none is set — skip to the next tier.
        return None

    try:
        return fetcher(api_key)
    except Exception:
        # Network error, auth failure, unexpected payload — degrade quietly.
        return None


def _fetch_openai(api_key: str | None) -> list[dict[str, Any]]:
    return _fetch_openai_compatible("https://api.openai.com/v1", api_key)


def _fetch_groq(api_key: str | None) -> list[dict[str, Any]]:
    return _fetch_openai_compatible("https://api.groq.com/openai/v1", api_key)


def _fetch_cerebras(api_key: str | None) -> list[dict[str, Any]]:
    return _fetch_openai_compatible("https://api.cerebras.ai/v1", api_key)


def _fetch_openai_compatible(
    base_url: str, api_key: str | None
) -> list[dict[str, Any]]:
    """Parse an OpenAI-shaped ``GET /models`` response."""
    data = _http_get_json(
        f"{base_url}/models",
        headers={"Authorization": f"Bearer {api_key}"},
    )
    entries: list[dict[str, Any]] = []
    for item in data.get("data", []):
        model_id = item.get("id")
        if not model_id or not _is_chat_model(model_id) or _is_fine_tune(model_id):
            continue
        created = _as_float(item.get("created"))
        entries.append(_entry(model_id, _humanize(model_id), created=created))
    return entries


def _fetch_anthropic(api_key: str | None) -> list[dict[str, Any]]:
    data = _http_get_json(
        "https://api.anthropic.com/v1/models",
        headers={"x-api-key": api_key or "", "anthropic-version": "2023-06-01"},
    )
    entries: list[dict[str, Any]] = []
    for item in data.get("data", []):
        model_id = item.get("id")
        if not model_id:
            continue
        label = item.get("display_name") or _humanize(model_id)
        created = _parse_iso(item.get("created_at"))
        entries.append(_entry(model_id, label, created=created))
    return entries


def _fetch_gemini(api_key: str | None) -> list[dict[str, Any]]:
    entries: list[dict[str, Any]] = []
    params: dict[str, Any] = {"key": api_key or "", "pageSize": 200}
    # models.list is paginated and not guaranteed newest-first, so walk pages
    # (bounded) to see the full set — _finalize does the sort + truncation.
    for _ in range(10):
        try:
            data = _http_get_json(
                "https://generativelanguage.googleapis.com/v1beta/models",
                params=params,
            )
        except Exception:
            # Later-page failure: keep the models already gathered. First-page
            # failure (nothing gathered yet) is a real outage — re-raise so the
            # caller falls back to the curated list rather than mistaking it for
            # a successful empty result.
            if entries:
                break
            raise
        for item in data.get("models", []):
            methods = item.get("supportedGenerationMethods") or []
            if "generateContent" not in methods:
                continue
            name = (item.get("name") or "").removeprefix("models/")
            if not name or not _is_chat_model(name) or "aqa" in name:
                continue
            label = item.get("displayName") or _humanize(name)
            # Gemini has no timestamp; rank by the version in name/version.
            version_hint = f"{name} {item.get('version') or ''}"
            entries.append(_entry(name, label, version_hint=version_hint))
        token = data.get("nextPageToken")
        if not token:
            break
        params = {"key": api_key or "", "pageSize": 200, "pageToken": token}
    return entries


def _ollama_base() -> str:
    """Resolve the Ollama server base URL from the environment.

    Checks ``OLLAMA_API_BASE`` / ``API_BASE`` (what LiteLLM and the generated
    crew use) first, then ``OLLAMA_HOST`` (the Ollama runtime convention), so a
    user who only set ``OLLAMA_HOST`` sees models from the right server.
    """
    base = (
        os.environ.get("OLLAMA_API_BASE")
        or os.environ.get("API_BASE")
        or os.environ.get("OLLAMA_HOST")
        or "http://localhost:11434"
    ).strip()
    # OLLAMA_HOST is often scheme-less (e.g. "127.0.0.1:11434").
    if "://" not in base:
        base = f"http://{base}"
    return base.rstrip("/")


def _fetch_ollama(_api_key: str | None) -> list[dict[str, Any]]:
    """List models installed on the local Ollama server (no API key)."""
    data = _http_get_json(f"{_ollama_base()}/api/tags")
    entries: list[dict[str, Any]] = []
    for item in data.get("models", []):
        model_id = item.get("model") or item.get("name")
        if not model_id or not _is_chat_model(model_id) or _is_fine_tune(model_id):
            # /api/tags lists everything installed, including embedding models.
            continue
        # Ollama returns an ISO 8601 modified_at we can rank by.
        created = _parse_iso(item.get("modified_at"))
        entries.append(_entry(model_id, _humanize(model_id), created=created))
    return entries


_VENDOR_FETCHERS: dict[str, Callable[[str | None], list[dict[str, Any]]]] = {
    "openai": _fetch_openai,
    "anthropic": _fetch_anthropic,
    "gemini": _fetch_gemini,
    "groq": _fetch_groq,
    "cerebras": _fetch_cerebras,
    "ollama": _fetch_ollama,
}


# ── Tier 2: LiteLLM feed ─────────────────────────────────────────

# Process-level memo so a single CLI run attempts the LiteLLM download at most
# once — repeated picker calls otherwise each incur a multi-second timeout when
# the feed is stale/unreachable. Reset via _reset_litellm_memo() in tests.
_UNSET: Any = object()
_litellm_memo: Any = _UNSET


def _reset_litellm_memo() -> None:
    """Clear the process-level LiteLLM memo (test hook)."""
    global _litellm_memo
    _litellm_memo = _UNSET


def _from_litellm(provider_key: str) -> list[dict[str, Any]] | None:
    """Build chat-model entries for ``provider_key`` from the LiteLLM feed."""
    data = _load_litellm_data()
    # A corrupt feed (non-mapping JSON root) must not crash the picker.
    if not isinstance(data, dict):
        return None

    entries: list[dict[str, Any]] = []
    for model_name, props in data.items():
        if not isinstance(props, dict):
            continue
        # `litellm_provider` can be present-but-null in the feed; coerce before
        # string ops so a null value is skipped rather than raising.
        if (props.get("litellm_provider") or "").strip().lower() != provider_key:
            continue
        if props.get("mode") != "chat":
            continue
        # LiteLLM keys are sometimes prefixed with the provider; the picker
        # re-adds ``provider/`` itself, so strip a leading one to avoid dupes.
        model_id = model_name
        if model_id.startswith(f"{provider_key}/"):
            model_id = model_id[len(provider_key) + 1 :]
        if not model_id:
            continue
        entries.append(_entry(model_id, _humanize(model_id), version_hint=model_id))
    return entries or None


def _load_litellm_data() -> dict[str, Any] | None:
    """Return the LiteLLM feed, memoized once per process (see _litellm_memo)."""
    global _litellm_memo
    if _litellm_memo is _UNSET:
        _litellm_memo = _fetch_litellm_data()
    memoized: dict[str, Any] | None = _litellm_memo
    return memoized


def _fetch_litellm_data() -> dict[str, Any] | None:
    """Read the cached LiteLLM feed, fetching it once if the cache is cold."""
    cache_file = _litellm_cache_file()
    fresh = (
        cache_file.exists()
        and (time.time() - cache_file.stat().st_mtime) < _LITELLM_TTL
    )
    if fresh:
        data = _read_json(cache_file)
        # A corrupt/non-mapping fresh cache must not block a recoverable
        # download — only short-circuit on a usable mapping.
        if isinstance(data, dict) and data:
            return data

    try:
        data = _http_get_json(JSON_URL)
    except Exception:
        # Fall back to a stale cache if we have one, else give up on this tier.
        return _read_json(cache_file)

    # Best-effort cache write; a failure (e.g. read-only home) is non-fatal
    # since we already hold the freshly-fetched data.
    with contextlib.suppress(OSError):
        cache_file.parent.mkdir(parents=True, exist_ok=True)
        cache_file.write_text(json.dumps(data), encoding="utf-8")
    return data


# ── Ranking + labelling ──────────────────────────────────────────


def _finalize(
    entries: list[dict[str, Any]], label_map: dict[str, str]
) -> list[tuple[str, str]]:
    """Sort newest-first, dedupe, relabel with curated names, and truncate."""
    entries.sort(key=lambda e: e["sort"], reverse=True)
    seen: set[str] = set()
    out: list[tuple[str, str]] = []
    for entry in entries:
        model_id = entry["id"]
        if model_id in seen:
            continue
        seen.add(model_id)
        label = label_map.get(model_id) or entry["label"]
        out.append((model_id, label))
        if len(out) >= MAX_MODELS:
            break
    return out


def _entry(
    model_id: str,
    label: str,
    *,
    created: float = 0.0,
    version_hint: str | None = None,
) -> dict[str, Any]:
    """Build a rankable catalog entry.

    ``sort`` is a comparable tuple ``(created, date_int, version_tuple)`` so a
    real vendor timestamp wins, then a date embedded in the id, then the numeric
    version. Types line up positionally, so entries compare cleanly.
    """
    date_int, version = _version_key(version_hint or model_id)
    return {
        "id": model_id,
        "label": label,
        "sort": (created, date_int, version),
    }


_DATE_RE = re.compile(r"(20\d{2})[-_]?(0[1-9]|1[0-2])[-_]?(0[1-9]|[12]\d|3[01])")
_NUM_RE = re.compile(r"\d+")


def _version_key(text: str) -> tuple[int, tuple[int, ...]]:
    """Extract a ``(date_int, version_tuple)`` sort key from a model id.

    A trailing/embedded ``YYYYMMDD`` (or ``YYYY-MM-DD``) becomes ``date_int``;
    remaining numbers become the version tuple. ``claude-opus-4-6`` → version
    ``(4, 6)``; ``claude-3-5-sonnet-20241022`` → date ``20241022`` version
    ``(3, 5)``.
    """
    text = text or ""
    date_int = 0
    match = _DATE_RE.search(text)
    if match:
        date_int = int(match.group(1) + match.group(2) + match.group(3))
        text = _DATE_RE.sub(" ", text)
    version = tuple(int(n) for n in _NUM_RE.findall(text)[:4])
    return date_int, version


def _is_chat_model(model_id: str) -> bool:
    """Heuristically reject embedding/audio/image/etc. models by their id."""
    lowered = model_id.lower()
    return not any(marker in lowered for marker in _NON_CHAT_MARKERS)


def _is_fine_tune(model_id: str) -> bool:
    """A user fine-tune or training checkpoint (``ft:...`` / ``...:ckpt-step-N``).

    These are account-specific artifacts: they clutter the picker, crowd out the
    foundation models (their recent ``created`` timestamps rank them first), and
    humanize into unreadable labels. Excluded from the auto-list; a user who
    wants one can still enter it via the picker's "Other" option.
    """
    lowered = model_id.lower()
    return lowered.startswith("ft:") or ":ckpt" in lowered


_SIZE_RE = re.compile(r"^\d+(?:\.\d+)?[bmk]$")  # 8b, 70b, 1.5b, 120m, 32k
_OSERIES_RE = re.compile(r"^o\d+$")  # o1, o3, o4 — kept lowercase (OpenAI brand)


def _humanize(model_id: str) -> str:
    """Derive a readable label from a raw model id.

    Best-effort only — vendor display names and the curated label map take
    precedence. Drops embedded dates and applies light casing so raw ids read
    cleanly: ``gpt-oss-120b`` → ``GPT OSS 120B``, ``qwen3-32b`` → ``Qwen3 32B``,
    ``deepseek-r1:671b`` → ``DeepSeek R1 671B``, ``o3-mini`` → ``o3 Mini``.
    """
    base = model_id.split("/")[-1]
    # Drop embedded release dates — they're noise in a label, and the picker
    # already shows the full model id alongside it.
    base = _DATE_RE.sub(" ", base)
    words: list[str] = []
    # Split on separators including ``:`` so Ollama tags (llama3.3:70b) read well.
    for part in re.split(r"[-_\s:]+", base):
        if not part:
            continue
        low = part.lower()
        if low in _ACRONYMS:
            words.append(_ACRONYMS[low])
        elif low in _BRAND_TOKENS:
            words.append(_BRAND_TOKENS[low])
        elif _SIZE_RE.match(low):
            words.append(low[:-1] + low[-1].upper())  # 70b -> 70B
        elif _OSERIES_RE.match(low):
            words.append(low)  # o3 stays lowercase
        elif part[0].isalpha():
            # Capitalize the leading letter, preserve the rest (so a fused
            # family+version keeps its digits): qwen3 -> Qwen3, mini -> Mini.
            words.append(part[0].upper() + part[1:])
        else:
            words.append(part)  # starts with a digit (4o, 4.1, 0905) — leave as-is
    return " ".join(words) or base


# ── HTTP + parsing helpers ───────────────────────────────────────


def _http_get_json(
    url: str,
    *,
    headers: dict[str, str] | None = None,
    params: dict[str, Any] | None = None,
) -> dict[str, Any]:
    """GET ``url`` and return parsed JSON, with a short timeout and TLS verify."""
    ssl_config = os.environ.get("SSL_CERT_FILE") or certifi.where()
    response = httpx.get(
        url,
        headers=headers,
        params=params,
        timeout=_TIMEOUT,
        verify=ssl_config,
        follow_redirects=True,
    )
    response.raise_for_status()
    result: dict[str, Any] = response.json()
    return result


def _parse_iso(value: Any) -> float:
    """Parse an ISO 8601 timestamp to an epoch float; ``0.0`` on failure."""
    if not value or not isinstance(value, str):
        return 0.0
    from datetime import datetime

    try:
        return datetime.fromisoformat(value.replace("Z", "+00:00")).timestamp()
    except ValueError:
        return 0.0


def _as_float(value: Any) -> float:
    try:
        return float(value)
    except (TypeError, ValueError):
        return 0.0


def _read_json(path: Path) -> dict[str, Any] | None:
    try:
        data: dict[str, Any] = json.loads(path.read_text(encoding="utf-8"))
        return data
    except (OSError, json.JSONDecodeError):
        return None


# ── Caching ──────────────────────────────────────────────────────


def _cache_dir() -> Path:
    return Path.home() / ".crewai"


def _catalog_cache_file() -> Path:
    return _cache_dir() / "model_catalog_cache.json"


def _litellm_cache_file() -> Path:
    # Shared with crewai_cli.provider so both flows warm the same cache.
    return _cache_dir() / "provider_cache.json"


def _is_cacheable(provider_key: str) -> bool:
    """Whether a provider's resolved catalog may be cached.

    Ollama is a local server (``/api/tags`` is fast), and its installed models
    change out-of-band, so it is never cached — the picker re-probes every call
    and always reflects what is currently installed.
    """
    return provider_key != "ollama"


def _cache_key(provider_key: str) -> str:
    """Cache key for a provider's resolved model list.

    Keyed by the exact API key (via a short, non-reversible digest — never the
    key itself), so switching to a different key for the same provider misses
    the previous account's cached entry and refetches. Absent key -> ``#nokey``,
    which also keeps a negatively-cached no-key fallback from shadowing a run
    after a key is added.
    """
    api_key = _provider_api_key(provider_key)
    if not api_key:
        return f"{provider_key}#nokey"
    digest = hashlib.sha256(api_key.encode("utf-8")).hexdigest()[:12]
    return f"{provider_key}#{digest}"


def _read_catalog_cache(provider_key: str) -> list[tuple[str, str]] | None:
    """Return a fresh cached catalog for ``provider_key``, or ``None``."""
    if not _is_cacheable(provider_key):
        return None
    payload = _read_json(_catalog_cache_file())
    if not isinstance(payload, dict):
        return None
    entry = payload.get(_cache_key(provider_key))
    if not isinstance(entry, dict):
        return None
    # Fallback (negative) entries expire fast; dynamic ones live the full TTL.
    ttl = _NEGATIVE_TTL if entry.get("source") == "fallback" else _CATALOG_TTL
    if (time.time() - _as_float(entry.get("ts"))) >= ttl:
        return None
    models = entry.get("models")
    if not isinstance(models, list) or not models:
        return None
    try:
        return [(str(m[0]), str(m[1])) for m in models]
    except (IndexError, TypeError):
        return None


def _write_catalog_cache(
    provider_key: str, models: list[tuple[str, str]], *, source: str
) -> None:
    if not _is_cacheable(provider_key):
        return
    cache_file = _catalog_cache_file()
    payload = _read_json(cache_file)
    if not isinstance(payload, dict):
        payload = {}
    payload[_cache_key(provider_key)] = {
        "ts": time.time(),
        "source": source,
        "models": [[model_id, label] for model_id, label in models],
    }
    # Best-effort cache write; a failure (e.g. read-only home) is non-fatal.
    with contextlib.suppress(OSError):
        cache_file.parent.mkdir(parents=True, exist_ok=True)
        cache_file.write_text(json.dumps(payload), encoding="utf-8")