dinos
83 lines
3.3 KiB
Python
83 lines
3.3 KiB
Python
"""Custom HTTP routes mounted alongside the langgraph dev server.
|
|
|
|
The langgraph-api host supports a top-level ``http`` key in
|
|
``langgraph.json`` that names an ASGI app to mount on the same
|
|
process as the graph. We use it to surface the registry the WebUI's
|
|
``/model`` picker needs.
|
|
|
|
Why this lives here and not as a separate sidecar: the WebUI talks to
|
|
``EvoSci deploy``'s langgraph endpoint anyway, so one origin keeps the
|
|
WebUI's fetch logic simple — no CORS dance, no extra port to configure.
|
|
|
|
Why Starlette and not FastAPI: ``langgraph_api`` already depends on
|
|
Starlette; adding FastAPI would pull in pydantic v1-vs-v2 reconciliation
|
|
the deploy doesn't need. The one route here has no input model, just a
|
|
JSON body, so the lower-level surface is sufficient.
|
|
|
|
Lightweight by design — module-level imports stick to ``config``,
|
|
``llm.models`` (registry only; no chat-model construction), and
|
|
Starlette itself. Nothing on this surface should pull the agent into
|
|
memory.
|
|
"""
|
|
|
|
from __future__ import annotations
|
|
|
|
import asyncio
|
|
|
|
from starlette.applications import Starlette
|
|
from starlette.requests import Request
|
|
from starlette.responses import JSONResponse
|
|
from starlette.routing import Route
|
|
|
|
from EvoScientist.config import get_effective_config
|
|
from EvoScientist.llm.models import list_model_picker_entries
|
|
|
|
|
|
async def get_models(_request: Request) -> JSONResponse:
|
|
"""Return the model registry as ``{entries, default}``.
|
|
|
|
``entries`` preserves the registry order so the WebUI picker can
|
|
rank providers per short name the same way the backend would.
|
|
Mirrors the TUI ``/model`` picker by appending locally-pulled
|
|
Ollama models when ``ollama_base_url`` is configured — same
|
|
``discover_ollama_models()`` call, same 1.5-s timeout, same
|
|
fail-soft semantics (the probe returns ``[]`` on any error, never
|
|
raises). The TUI's "Custom Ollama model…" sentinel is intentionally
|
|
omitted: that's a widget-specific input affordance, not part of
|
|
the registry surface.
|
|
|
|
``default`` reflects the deployment's currently-configured fallback
|
|
(``config.yaml``'s ``model`` / ``provider`` — what ``/model reset``
|
|
would land on). Returned even when the configured pair isn't in
|
|
the registry, so the picker can still label it.
|
|
|
|
Uses ``get_effective_config()`` (not ``load_config()``) so env-var
|
|
overrides like ``OLLAMA_BASE_URL`` from ``_ENV_MAPPINGS`` are
|
|
honored — matching the deploy's actual model-building behavior.
|
|
Offloaded to a thread because ``get_effective_config()`` calls
|
|
``find_dotenv(usecwd=True)`` which invokes ``os.getcwd()`` — a
|
|
blocking syscall that langgraph-dev's ``blockbuster`` middleware
|
|
refuses to allow on the async event loop (would surface as a 500).
|
|
"""
|
|
cfg = await asyncio.to_thread(get_effective_config)
|
|
entries = [
|
|
{"name": name, "model_id": model_id, "provider": provider}
|
|
for name, model_id, provider in await list_model_picker_entries(
|
|
getattr(cfg, "ollama_base_url", None),
|
|
include_custom_ollama=False,
|
|
)
|
|
]
|
|
return JSONResponse(
|
|
{
|
|
"entries": entries,
|
|
"default": {"name": cfg.model, "provider": cfg.provider},
|
|
}
|
|
)
|
|
|
|
|
|
app = Starlette(
|
|
routes=[
|
|
Route("/api/models", get_models, methods=["GET"]),
|
|
]
|
|
)
|