Repository navigation
To upgrade run the upgrade_packages.sql script attached to the GitHub release. This release has no migration script.
For new installations follow this guide.
Behavior notes:
- TOON output follows specification 4.1. The
uc_ai_toonencoder now writes an empty array askey: []and no longer askey[0]:. It puts a key outside letters, digits,_and.in quotes, and a string that starts with#or-. A column that holds objects with the same structure folds into the header, and an object whose values share the same structure uses a keyed table. The encoded text of some inputs is therefore different from earlier releases. A test that compares TOON text exactly needs new expected values. See TOON.
Features:
OCR:
- OCR (
uc_ai.ocr,uc_ai.ocr_text): extract the text of a PDF or an image, without a chat model. The result has the same shape for every provider: the text as Markdown, the pages, the text blocks with normalized boxes and confidence where the provider reports them, the usage, the warnings, and the response of the provider. Three providers read documents: Mistral (the OCR modelsuc_ai_mistral.c_model_mistral_ocrandc_model_mistral_ocr_4_1, PDF, images, DOCX and PPTX, and a URL overload), OCI Document Understanding (PDF, PNG, JPEG and TIFF, with table extraction), and Ollama (images, read by a vision model that you name). The neutral optionspagesandtablesand the keyextra_bodywork across providers. See the provider pages for Mistral, OCI and Ollama.
New models:
- New model constants. OpenAI:
c_model_gpt_6_1_sol,c_model_gpt_6_sol, andc_model_gpt_6_luna. Anthropic:c_model_claude_5_5_opusandc_model_claude_5_5_sonnet. xAI:c_model_grok_4_7. Mistral:c_model_zai_glm_5_3. OCI:c_model_grok_4_7andc_model_grok_4_6.
Tutorials:
- Answer from Your Documents: a new course that builds retrieval-augmented generation (RAG) on your own documents.
Fixes:
- Mistral chat models read a PDF. UC AI sends a PDF in a message to Mistral as a
document_urlcontent part. See File analysis.
Full Changelog: v26.3...v26.4