Back to the room

Pass 3 of 3 · Deep dive of PHAROSPass 1 · OverviewPass 2 · How it works

Code Observatory / class

EdgeFirstEmbedder

codeintel.embeddings.EdgeFirstEmbedder

PINNED SOURCE Source and edges from the index at this commit. An execution appears only where one was recorded.

PHAROS at 975a38dMIT license

Around this symbol

Search all symbols

Loading

Called by (1)
SymbolFileResolved by
get_embedderembeddings.pysame file
Contains (8)
SymbolFileResolved by
__init__embeddings.pysame file
_edgeembeddings.pysame file
_httpembeddings.pysame file
_reset_httpembeddings.pysame file
describeembeddings.pysame file
embed_documentsembeddings.pysame file
embed_queryembeddings.pysame file
encode_rawembeddings.pysame file
Inside (1)
SymbolFileResolved by
embeddingsembeddings.pysame file
codeintel.embeddings.EdgeFirstEmbedderbackend/app/codeintel/embeddings.py, lines 367–426Open source
class EdgeFirstEmbedder(Embedder):    """The edge's embed server first, the local ONNX model as the fallback.    Both run the same model, so their vectors are interchangeable; the edge's    ``model_id`` is checked on every answer and a different model is never    used. The local model is loaded only when the edge fails (timeout, error,    wrong model), and after a failure the edge is skipped for ``retry_after``    seconds so a sleeping edge does not add its timeout to every query.    """    def __init__(self, base_url: str, token: str, local: OnnxEmbedder, *, timeout: float = 2.5,                 retry_after: float = 60.0):        self.base_url = base_url.rstrip("/")        self.local = local        self.model_id = local.model_id        self.timeout = timeout        self.retry_after = retry_after        self._headers = {"Authorization": f"Bearer {token}"}        self._client = None  # one keep-alive connection: a TLS handshake per query cost ~0.5 s through Funnel        self._client_lock = threading.Lock()        self._skip_until = 0.0        self.edge_calls = self.local_calls = 0    def _http(self):        import httpx        with self._client_lock:            if self._client is None:                self._client = httpx.Client(base_url=self.base_url, headers=self._headers, timeout=self.timeout,                                            limits=httpx.Limits(max_keepalive_connections=2, keepalive_expiry=300))            return self._client    def _reset_http(self) -> None:        with self._client_lock:            client, self._client = self._client, None        if client is not None:            client.close()    def _edge(self, texts: list[str]) -> list[list[float]]:        import httpx        try:            resp = self._http().post("/embed/batch", json={"texts": texts})        except httpx.TransportError:            # An idle connection the edge or the proxy closed: retry once on a            # fresh one before falling back (the fallback loads the local model).            self._reset_http()            try:                resp = self._http().post("/embed/batch", json={"texts": texts})            except Exception:                self._reset_http()                raise        if resp.status_code != 200:            raise EmbeddingUnavailable(f"edge returned HTTP {resp.status_code}")        data = resp.json()        if data.get("model_id") != self.model_id:            raise EmbeddingUnavailable(f"edge runs {data.get('model_id')!r}, not {self.model_id!r}")        vectors = data["embeddings"]        if len(vectors) != len(texts):            raise EmbeddingUnavailable("edge returned the wrong number of vectors")

Shows the first 60 of 84 lines. The rest is on GitHub at this commit.

What the index recorded

Signature
class EdgeFirstEmbedder(Embedder):
File
backend/app/codeintel/embeddings.py
Lines
367–450
Parser
python-ast
Git blob
df7f8202f23ca6ad2c57aca5f09b435eb6486d79
Outside this subset
6 more edges reach code outside the graph shown here

Read it on GitHub at this commit