Document AI 1.0.0 Release Notes

Document AI 1.0.0 is the first general release of Document AI.

Document AI is the shared document-intelligence service of the C3 Agentic AI Platform. It turns large volumes of unstructured documents, such as contracts, reports, manuals, and filings, into governed application context. Applications and agents can then retrieve and reason over that structured, searchable content.

Document AI processes documents as a shared platform service, returning ready-to-use results to any application that calls it. An application submits one or more documents together with a processing strategy. Document AI parses, chunks, and embeds each file, and tracks every file through to completion. It can then write the results back to the application as searchable vector passages or entities. Each job can be followed to completion, with logs, token usage, and cost available for review.

For a product tour, see Document AI Overview. To process your first document in about ten minutes, see Document AI Quick Start.

Copy link to this section

Key features

Document AI 1.0.0 introduces the following capabilities:

• Document Processing: Document AI can parse, chunk, and embed one or more documents in a single submission, producing content that is ready for search and retrieval-augmented generation. Supported formats include PDF, Office documents (Word .docx/.doc, PowerPoint .pptx/.ppt, Excel .xlsx/.xls), plain text formats (.txt, .csv, .html, .rtf), and images (.jpeg, .jpg, .png).

• Processing Strategies: Choose a strategy to control how Document AI handles each document. Passage-level splits the document into chunks and embeds them for search. Document-level extracts whole-document metadata without chunking or embedding. Teams can also supply a custom strategy inline as Python source on each request, with no registration step and no redeployment; see Develop Custom Strategies.

• REST API: A documented REST API lets any C3 application submit documents and persist results. Applications call the API directly from their own Console, without needing direct access to the Document AI application itself. For prerequisites, endpoints, and examples, see Document AI REST API Developer Guide.

• Agentix Assistant: Ask the Agentix assistant to process documents in plain language. For example, "Ingest report.pdf into the default-pg retriever." The assistant submits the job, monitors it to completion, and reports the outcome, including for files attached directly in chat.

• Jobs and Runs: Each submission becomes a job that tracks every file individually. At a glance, you can see what has completed, what is still processing, and what has failed. Results can then be retrieved per file.

• Automatic Vector Persistence: Point a job at a retriever, and Document AI automatically writes successful results back to your application as searchable vector passages, ready for retrieval-augmented generation. You can also persist a completed run manually at any time. Do this for jobs submitted without a retriever, or to retry after a persistence error. Persisting the same document into the same retriever again is safe, because Document AI replaces stale passages instead of duplicating them.

• Tagging and Entity Extraction: Beyond text passages, Document AI can use the language model to pull structured tags and named entities out of each document. It writes them back to your application, as tags on the source file and as records in your data model. Extraction runs only for the tags and entities you request, so passage-only workloads add no extra cost. For schema configuration, see Configure Tag and Entity Extraction.

• Size-Tiered Processing Fleets: Document AI routes each file to a small, medium, or large processing fleet by file size so a small file never waits behind a large one. Default processing fleet sizes:

  • Files up to 50 MB use the small fleet
  • Files up to 200 MB use the medium fleet
  • Files up to 1 GiB use the large fleet
  • Files above 1 GiB are rejected at submission

• Autoscaling: Each size-tier fleet scales independently from the depth of its own work queue. The small fleet stays available with a minimum of one pod by default. The medium and large fleets scale from zero when work arrives and scale back to zero when idle. The first medium or large file after an idle period may wait briefly while that fleet scales up.

• Reliability Controls: Built-in concurrency limits, automatic retries, and per-file size limits keep large ingestion jobs stable and predictable, and Document AI reports every failure. Administrators can tune these limits to match their workload:

  • Retry count: 3 by default.
  • Processing timeout: 120 minutes by default.
  • Per-file size limit: 1 GiB by default (also the large-tier ceiling).

• Governed Model Access: All language model and embedding calls route through the C3 Agentic AI Platform's centralized model gateway: the single point of governance every platform service shares. The processing service stores no model provider credentials. Document AI automatically grants every authenticated user the DocumentAi.User role, which covers all processing endpoints, eliminating the need for per-user setup. Viewing token consumption and cost on the usage dashboard requires the separate DocumentAi.Admin role.

• Per-Job Model Overrides: Override the completion model, the embedding model, or both for a single processing job, without changing the service's default configuration or affecting any other job. Use a model override to compare model quality or cost, or to embed a corpus for a retriever built on a different embedding model. Document AI validates the override before creating any run, and records the models used on the job so you can attribute usage and cost correctly.

Copy link to this section

Compatibility

Document AI 1.0.0 runs on the C3 Agentic AI Platform 8.10 or later. The LLM service it depends on is installed alongside it.