Skip to content

AGENTS.md

Latest commit

 

History

History
45 lines (35 loc) · 1.62 KB

AGENTS.md

File metadata and controls

45 lines (35 loc) · 1.62 KB
Raw

AGENTS.md

Purpose

@corbits/embedding turns texts into vectors over any OpenAI-compatible /v1/embeddings endpoint, with @intx/inference retries and error classification. It owns batching, reply integrity and the one-shot JSON transport. It does not store or search vectors; that is @corbits/memory.

Layout

  • src/embed.ts: EmbedConfigSchema, embedTexts, probeEmbedDims, batch splitting and reply parsing (float and base64).
  • src/request.ts: runJSONRequest (private transport), EmbeddingRequestError, Retry-After extraction.
  • src/index.ts: the only module consumers import from.
  • e2e/: harness-driven retry, wire and base64 suites, plus a live round trip gated on EMBEDDING_E2E_BASE_URL.

Rules

  • Place vectors by the echoed index, never by array position. Reject short, duplicate or out-of-range replies instead of mispairing.
  • Batches run sequentially; retry policy is per request.
  • Throw EmbeddingRequestError carrying the classified InferenceError as reason; never throw the InferenceError itself.
  • Read a 200 as text, then JSON.parse, so a non-JSON body classifies as a protocol mismatch.
  • Omit dimensions and encoding_format from the wire unless configured.
  • Backoff time comes from deps.scheduler, not wall-clock globals.
  • exactOptionalPropertyTypes is on: omit optional keys, never assign undefined to them.

Local development

bun install
bun run check

Conventions

  • Read README.md before changing public behavior.
  • Keep changes focused; update docs and tests alongside code changes.
  • Run bun run check before declaring work complete.