Changelog

vNext

Add support for translation endpoint i.e. whisper models. (#59 @wtulloch)
- This also includes internal changes to support non-token based rate limiting
Add support for specifying the dimension parameter in in embeddings requests for text-embedding-3 and later models (#55 - @tanya-borisova)
Ensure that an API key is always generated if not provided (#56 - @lucashuet93)
BREAKING CHANGE Requests for an incompatible model (e.g. chat requests for an embedding model) fail with a 400 error (#58 - @tanya-borisova)
Terraform deployment option (#60 @mluker)
Support for ARM architecture for local Docker builds (#32 @mluker)
Numerous fixes and repo improvements: #24, #26, #38, #41, #42, #43, #45, #51 @martinpeck

Migrate to current repo from previous repo
- BREAKING CHANGE:: rename aoai-simulated-api to aoai-api-simulator in code (also for aoai_simulated_api package)
- BREAKING CHANGE:: update metric prefix from aoai-simulated-api. to aoai-api-simulator.
Return to sliding window rate limiting. This change moves from the limits package to a custom rate-limiting implementation to address performance with sliding windows (#51)
Update rate-limit handling for tokens based on experimentation (limited set of models currently - see #52)

Extensibility updates
- Focus core simulator on OpenAI (moved doc intelligence generator to example extension)
- API authorization is now part of forwarders/generators to allow extensions to add their own authentication schemes. BREAKING CHANGE: If you have custom forwarders/generators they need to be updated to handle this (see examples for implementation details)
- Enable adding custom rate limiters
- Move latency calculation to generators. This allows for extensions to customise latency values. NOTE: If you have custom generators they need to be updated to handle this (see examples for implementation details)
Add rate-limiting for replayed requests
Add ALLOW_UNDEFINED_OPENAI_DEPLOYMENTS configuration option to control whether the simulator will generate responses for any deployment or only known deployments
Fix: tokens used by streaming completions were not included in token counts for rate-limits
Token usage metrics are now split into prompt and completion tokens using metric dimensions
BREAKING CHANGE: Token metrics have been renamed from aoai-simulator.tokens_used and aoai-simulator.tokens_requested to aoai-simulator.tokens.used and aoai-simulator.tokens.requested for consistency with latency metric names
Dimension size for embedding deployments can now be specified in config (#39 - @MDUYN)

Initial tagged version

Includes

Update Dockerfile to enable building with pre-cached tiktoken file (enables running without internet access) (#23 @aerjenn)
Document Intelligence example (#2 @mcollier)