Skip to content

CHANGELOG.md

Latest commit

 

History

History
185 lines (154 loc) · 7.09 KB

CHANGELOG.md

File metadata and controls

185 lines (154 loc) · 7.09 KB
Raw

Typeagent Change Log

2026

0.5.0 (May 26)

This release focuses on ingestion throughput, robustness, and developer tooling, with a new streaming ingestion pipeline and substantial performance work. Thanks go out to:

  • Bernhard Merkle
  • Kevin Turcios
  • Shreejay Kurhade
  • Dhaval Gojiya
  • Tushar Mudi
  • ANGELUSD11
  • dependabot[bot]

Core typeagent package

  • Deliver a new high-throughput streaming ingestion architecture and a new API add_messages_streaming, with stronger correctness and broader test coverage (#252, #266, #277).
  • Move ingestion workflows onto the streaming path (notably email and VTT), simplifying duplicate filtering logic (#265, #267, #268, #271).
  • Tune embedding model parameters per-model and add supporting benchmark (#228).
  • Improve ingestion and query performance via batched SQLite writes, batched metadata lookups (removing N+1 paths), overlapped streaming pipeline stages, numpy-vectorized embedding lookup, and higher podcast ingest concurrency (#230, #232, #234, #248, #260).
  • Replace ad-hoc retries with stamina and increase retries to better handle 429s (#255, #245).
  • Resolve Azure endpoint/URL parsing issues and regressions, with expanded parsing tests (#231, #239, #241, #244).
  • Turn async provider accessor methods into sync properties for provider attributes (#256).
  • Remove dead and obsolete internal code paths (extract_knowledge_for_text_batch_q, create_async_openai_client) (#254, #259).
  • Rename kplib.py to knowledge_schema.py (#212).
  • Remove runtime black dependency: switch to using pprint (#229, #235, #237).

Tools

  • Add benchmark scripts under tools/ (benchmark_query.py, benchmark_vectorbase.py, benchmark_embeddings.py, repeat_embedding_benchmarks.py) (#251, #228, #247).
  • Improve streaming ingestion tool ergonomics in ingest_email.py and ingest_vtt.py with explicit batching/concurrency controls and clearer per-batch progress reporting (#265, #267, #268).
  • Add utility to convert mbox files to .eml files (#220).
  • Add new chat_sessions.py tool to browse/dump VS Code Copilot chat sessions (#226, #233).
  • Add make eval target (#237).

Docs

  • Document high-level ingestion methods in API docs (#261).

Testing and coverage

  • Improved test coverage -- somewhat.
  • Fixed coverage support on Windows (.coveragerc, make.bat) (#249, #253).

CI and security

  • Harden CI workflow permissions and trigger model (pull_request, remove unnecessary id-token: write) (#214).
  • Apply workflow security fixes reported by zizmor across CI/CodeQL and release
  • workflows (#276).

0.4.0 (March 3)

Lots of improvements; the highlights are provider-agnostic model configuration backed by pydantic_ai, email ingestion, and a major infrastructure overhaul.

Core typeagent package

  • Fixed a number of bugs that affected the core query algorithms
  • Provider-agnostic chat and embedding model configuration via new model_adapters module backed by pydantic_ai (#200):
    • Use provider:model spec strings, e.g. create_chat_model("openai:gpt-4o").
    • Replace AsyncEmbeddingModel with IEmbedder/IEmbeddingModel protocols and CachingEmbeddingModel.
    • Add OPENAI_MODEL and OPENAI_EMBEDDING_MODEL envvars to override the default chat and embedding models.
  • Split interfaces.py into separate modules (interfaces_core, _indexes, _search, _serialization, _storage) (Bernhard Merkle, #118).
  • Make remaining storage-provider APIs async (get/set_conversation_metadata, is_source_ingested, etc.) (#196).
  • Fix listeners/recipients confusion in podcast metadata serialization (#174).
  • Implement SqliteRelatedTermsIndex.serialize() (Rajiv Singh, #115).

Email

  • New tools/ingest_email.py tool to ingest email into a SQLite-backed conversation database (#111).
  • Add tools/mail/outlook_dump.py to dump Outlook/Microsoft 365 email via the Graph API (Bernhard Merkle, #199).
  • Add tools/mail/mbox_dump.py to convert mbox files for ingestion (Bernhard Merkle, #198).
  • Consolidate mail dump tools under tools/mail/ with shared README.md (Bernhard Merkle).
  • Various ergonomic improvements and fixes (#162, #168, #170).

Tools

  • Add conversation history to tools/query.py for pronoun/reference resolution across multi-turn queries (Rajiv Singh, #117).
  • Add tools/load_json.py to load JSON-serialized index data into a SQLite database; remove --podcast flag from tools/query.py (#164).

Docs

  • Improve docs for Azure env vars (#175).
  • Add AgentCon 2026 presentation and demo videos (Bernhard Merkle, #194, #202).
  • VS Code / Pyright plugin setup instructions (Bernhard Merkle, #150).

Infrastructure

  • Changes pyproject.toml to use uv more idiomatically.
    • Local devs write uv sync instead of uv sync --extra dev.
    • From PyPI (with uv or pip) you can use typeagent[dev] to install the dev dependencies with the package.
  • Move typeagent/ to src/typeagent/ (Bernhard Merkle, #139).
  • Move tests and test data to tests/ directory (Bernhard Merkle, #144).
  • Move ancillary dirs into subdirs (Bernhard Merkle, #145).
  • Introduce isort for import sorting.
  • Make pyright error on unused variables and imports (#129).
  • Add readline support on Windows (#152, Bernhard Merkle).
  • Enhance release script to update uv.lock and create release PR (Rajiv Singh, #169).

2025

0.3.3 (Nov 25)

General cleanup and fixes, and the following notable improvements:

Docs

  • Document AZURE_OPENAI_ENDPOINT_EMBEDDING (Gwyneth Peña-Siguenza).
  • Add link to PyBay 2025 talk video.

Core typeagent package

  • Simplify load_dotenv().
  • Split embedding requests that are too large (Raphael Wirth).
  • Overhaul conversation metadata in storage providers.
  • Add extra optional keyword parameters to create_conversation().
  • Add Quantifier to ingestion schema in addition to Quantity.
  • Extract knowledge concurrently (max 4 by default) (Kevin Turcios).
  • Fixes for get_multiple() implementations.
  • Tweak defaults in ConversationSettings.

MCP server

  • Pass LLM requests to MCP client instead of calling the OpenAI API.
  • Add --database option to MCP server.

Tools

  • The tools/query.py tool now supports @-commands. Try @help.
  • Add tools/ingest_podcast.py (a tool that ingests podcasts).

Testing

  • Fix coverage support for MCP server test.
  • Use an updated "Adrian Tchaikovsky" podcast data dump (Rob Gruen).
  • Fix Windows testing issues. Run CI on Windows too (Raphael Wirth).
  • Run tests in CI using secrets stored in repo (Rob Gruen).

Infrastructure

  • Migrate package build from setuptools to uv_build.
  • Add install-libatomic target to Makefile (Bernhard Merkle).

0.3.2 (Oct 22)

Brown bag release!

  • Put black back with the runtime dependencies (it's used for debug output).

0.3.1 (Oct 22)

  • Limit dependencies to what's needed at runtime; dev dependencies can be installed separately with uv sync --extra dev.
  • Add endpoint_envvar arg to AsyncEmbeddingModel to allow configuring a non-standard embedding service.

0.3.0 (Oct 17)

  • First public release, for PyBay '25 talk