Skip to content

v3.5.0

Choose a tag to compare

View all tags
@timothymarois timothymarois released this 10 Jun 22:30
· 71 commits to 3.x since this release
v3.5.0
818a520

Quality of life improvement update

Added

  • New typed exceptions for specific provider failures: ConnectionException (network), ModelNotFoundException (unknown model), InvalidRequestException (bad request), and ServerException (provider error).
  • xAI now supports the CodeInterpreter provider tool.
  • voice.route_middleware config to protect the voice routes, e.g. ['auth:sanctum', 'throttle:60,1'].
  • ProviderException::responseBody() and rawResponse() expose the provider's raw error response for debugging, without unwrapping the previous exception.
  • Provider request events (ProviderRequestStarted / Completed / Failed / Retrying) now carry a correlationId (stable across retries) plus provider and model, so you can trace and attribute every HTTP call.

Changed

  • atlas.embeddings.dimensions is now sent to OpenAI's text-embedding-3-small/-large as the dimensions parameter, so reduced-dimension (Matryoshka) embeddings work straight from the config key. An explicit dimensions provider option still wins; models that don't support it (e.g. text-embedding-ada-002) are unaffected.
  • A single catch (ProviderException) now handles every provider failure. Catch a subclass for specific cases.
  • "Overloaded" provider responses are now retried, like rate limits.
  • Attaching an unsupported provider-native tool (e.g. xAI's X Search on OpenAI) now throws UnsupportedFeatureException up front instead of failing at the provider.

Fixed

  • Structured output via the Schema builder now works on OpenAI, xAI, and OpenAI-compatible providers (Ollama, LM Studio) — previously these 400'd because strict JSON-schema requirements (additionalProperties: false, all keys required, optionals as nullable) weren't applied. Optional fields now round-trip as nullable.
  • Rate limits and transient server errors now retry automatically.
  • Per-call and global timeouts are honored, including on queued requests.
  • Network failures and mid-stream errors now surface as exceptions instead of returning truncated, successful-looking responses.
  • Interrupted streams broadcast the usage and finish reason captured before the break, so cost tracking survives.
  • Queued requests fail fast on unrecoverable errors (bad key, bad request, unknown model) instead of burning every retry.
  • Per-request middleware (->withMiddleware()) now runs on queued image, audio, video, speech, embedding, moderation, rerank, music, and sound-effect requests — previously it was silently dropped on queue (text and agent requests were unaffected).
  • Queuing a request with a closure middleware now fails fast with a clear error instead of crashing the worker; use class-based middleware for queued requests.
  • Batch embeddings are realigned by the provider's index field rather than response order, so an out-of-order batch can't attach the wrong vector to a chunk.
  • A mismatched embedding dimension now fails at embed time with a clear message (actual vs configured size, plus the fix) instead of a cryptic database error at write time.
  • Changing config at runtime and calling AtlasConfig::refresh() now applies to facade calls (Atlas::embed(), etc.) — previously the facade held a stale manager built from the old config, so runtime provider/model/dimension changes were silently ignored on facade-driven paths (embeddings, chunking).
  • Google tool calls missing a function name now degrade gracefully instead of erroring mid-parse.
  • Listing models and voices now reports failures like any other call.
  • Error messages now carry the provider's real reason across all providers — including auth failures (401/403), which previously showed a generic message instead of the provider's text (e.g. "Incorrect API key provided").
  • Queued failures cap their broadcast error so it can't exceed the socket limit and drop the event.

Migration

No breaking changes for most users. Check these only if they apply:

  • You hard-coded a vector(3072) column for text-embedding-3-large but left ATLAS_EMBEDDING_DIMENSIONS at the 1536 default — that key is now forwarded to the API, so set it to 3072 to match your column. (Sizing the column from config('atlas.embeddings.dimensions') already handles this.)
  • You catch specific exceptions (auth, rate-limit) before ProviderException — keep those catch blocks first; ProviderException now catches them too.
  • You use reasoning_timeout — removed. Set a longer per-call timeout instead.
  • You read streamed responses — mid-stream errors now throw. Wrap your read loop in try/catch.
  • You attach provider-native tools to providers that don't support them — now throws UnsupportedFeatureException up front.

Pull Requests

Full Changelog: v3.4.0...v3.5.0

Contributors

timothymarois
Assets 2
Loading