0.3.0

Breaking changes

• DialException is materialized instead of being treated as a normal chat completion chunk (this is motivated by DIAL SDK not being able to handle error chunks properly). This led to ChatCompletionInterceptor.call_upstreams changing type signature to accommodate the new kind of stream that includes DialException values (#13):

  Before:

  class ChatCompletionInterceptor(RequestHandler, ResponseHandler):
      async def call_upstreams(
          self,
          request: dict,
          call_upstream: Callable[
              [dict, Any | None], Coroutine[Any, Any, AsyncIterator[dict]]
          ],
      ) -> AsyncIterator[AnnotatedChunk]:
      ...

  After:

  class ChatCompletionInterceptor(RequestHandler, ResponseHandler):
      async def call_upstreams(
          self,
          request: dict,
          call_upstream: Callable[
              [Any | None, dict], Awaitable[AsyncIterator[dict | DialException]],
          ],
      ) -> AsyncIterator[AnnotatedChunk | AnnotatedException]:
      ...

Fixes

• fixed propagation of exceptions from upstream endpoints (#13)
• fixed printing of debug logs (#13)
• fixed bug in algorithm that merges chat completion chunks in a non-streaming mode (#13)
• introduced X-UPSTREAMS header and refactored DIAL Client to ease interceptor testing (#13)
• examples[pii]: en_core_web_sm model is (a) downloaded as part of the package build and (b) pre-loaded on the server start-up instead of spending time of the first request into PII interceptor (#13)
• examples[pii]: pinned versions of en_core_web_sm model (3.7.1) and spacy (3.7.5) to avoid model downloading issue on the server start up (#13)
• examples[pii]: removed en_core_web_sm as a package dependency since PyPi doesn't support direct dependencies (#16)

Other

• add reference to PyPI (#14)
• bump aiohttp from 3.10.2 to 3.10.11 (#12)
• bump epam/ai-dial-ci from 1.9.1 to 1.10.0 (#7, #10, #11)
• bump protobuf from 5.29.0 to 5.29.1 (#17)
• bump starlette from 0.36.3 to 0.40.0; fastapi from 0.109.2 to 0.115.2 (#8)
• bumped github workflow versions (#6)
• optimize trivy (#15)
• built infrastructure for testing chains of interceptors (#13)
• added tests that validate response headers propagation from the upstream in streaming and non-streaming scenarios (#13)
• remove aiohttp from the package dependencies (#13)
• aiostream package dependency moved to the examples extras (#13)
• bump aidial-sdk from 0.13.0 to 0.16.0 (#13)