Changelog

The notable changes in each release. Every release on GitHub also carries auto-generated notes listing every merged pull request, and Upgrading covers the steps an existing deployment has to take.

0.20.0

DocsGPT installs from PyPI

The backend is published as docsgpt: the API server, the web UI, the Celery worker and the maintenance scripts in one package, behind a single docsgpt command. pip install docsgpt then docsgpt api serves the API and the UI on one port, and docsgpt worker runs the worker. The optional engines are extras, docsgpt[docling] and docsgpt[milvus]. See Install with pip.

The Python package is now `docsgpt`

The import package was renamed from application to docsgpt, the name it has on PyPI. Entry points move with it: celery -A docsgpt.app.celery worker and uvicorn docsgpt.asgi:asgi_app. The old spellings still run for this release and print a FutureWarning, and Celery tasks queued under the old names are still consumed. The upgrade guide has the details.

Smaller default install and images

The document and vector-store engines that pulled the heaviest dependencies are now extras rather than defaults, so a stock install no longer carries PyTorch or the CUDA stack. The published image follows the same split: the default arc53/docsgpt image is the slim one, and arc53/docsgpt:<version>-docling bakes in the docling parser engine, its models and tesseract for OCR.

Faster embeddings, pinned to your installation

Embeddings run through FastEmbed on ONNX Runtime. Query embedding now happens on the Celery worker rather than in the API process, which keeps the API small; a worker consuming the embeddings queue is required for search. The embedding model is pinned to the installation instead of following the release, so an upgrade never silently changes the model behind an existing index, and a query against an index built with a different model is now warned about. New installs default to ibm-granite/granite-embedding-311m-multilingual-r2, which is multilingual with a 32k-token context. The reembed script rebuilds vectors in place from the chunk text already stored.

If you start your worker with an explicit -Q, add the embeddings queue. See Upgrading for the one-line change and what happens without it.

Broader document support

Parsing gained support for more document, spreadsheet, presentation, EPUB, XHTML and image formats, along with a reworked OCR path. Parsing behaviour is configurable: markdown conversion, structured output, table reconstruction and improved PDF handling.

Agents and workflows

Agents no longer require a source. The synthetic “Default” source is gone, and an agent with no source or retriever is a valid agent that lists and answers normally. Workflow agents can be exported and imported as files, so a workflow can move between deployments or into version control.

Artifact and tool-call handling was hardened throughout: durable tasks retry rather than lose work on resume, tool calls are validated more strictly, and sandbox sessions are steadier.

Chat and attachments

Cross-turn chaining against the Responses API is now bounded, conversation compression persists across turns instead of being recomputed, and requests carry prompt-cache hints. Attachments carry provenance through parsing, unparseable chat attachments are refused at upload rather than failing mid-answer, and the composer guards against sending while an attachment is still processing.

The React widget got a UI refresh and an expand and collapse toggle, and was brought back in line with the current API contract. The npm packages docsgpt and docsgpt-react are published at 0.7.1.

Also in this release

The code-execution sandbox runner is published as arc53/docsgpt-sandbox, so the Kubernetes manifest no longer needs an image you build yourself.
Releases publish the Docker images and the PyPI package from the release workflow.
The architecture guide was rewritten.
Markdown code spans are no longer corrupted by citation rendering.
setup.ps1 runs on Windows PowerShell 5.1 again.