Starstruk Logo MarkStarstruk

v1.0.0-rc2 v0.27.3

high risk

infiniflow/ragflowOct 16, 2026 → Oct 23, 20267 days between releases

What's New

v0.27.3 ports Agentic RAG to Go and swaps Valkey for Kvrocks, with schema migrations and parser contract changes.

Breaking Changes

  • Cache/queue backend moves from Valkey 8 to apache/kvrocks:2.16.0, and the Go path now defaults to Kvrocks and fails fast if it is unreachable.
  • Chat messages and reference data are split into dedicated database tables, so an explicit schema migration is required before serving traffic.
  • Parser outputs are unified into structured JSON items, changing the parser/chunker contract for anything consuming parse results directly.
  • Internal CLI commands for inserting chunks and metadata from file are removed, along with Go artifact-deletion APIs and go-ingestion checkpoint/resume support.

New Features

  • Agentic RAG is reimplemented in Go, including a native Go MCP runtime and Go sandbox settings with live refresh.
  • New data-source connectors for Feishu Wiki (Python and Go syncers) and Zotero.
  • Additional model providers: AnonRouter, API-Route, DaoXE, Cheaper Inference, the OpenAI GPT-5.6 family, and recent Gemini models.
  • Reranking now operates on natural chunk text rather than tokenized content_ltks.
  • Claim/evidence support added to the tree/pageindex knowledge compilation pipeline.

Improvements

  • Broad security hardening across SSRF guards (IPv6 transition addresses, connector SSRF), dataset/document authorization, metadata filter scoping, and orderby allowlisting.
  • Ingestion pipeline refactored so parsing state is decoupled from document.run into ingestion_task with run-scoped progress events.
  • DeepDoc/ONNX inference tuned for better performance and lower memory use.
  • Cache/queue stack consolidated on Kvrocks 2.16.0 for both Python and Go services.
  • Removal of deprecated and unused code paths, including unused knowledge compilation code.

Bug Fixes

  • MCP tool inputSchema.properties are now translated correctly in Info().
  • rerank_candidates_count is sent so retrieval paging works beyond 64 results.
  • Agent retrieval knowledge bases are compared by resolved embedding model instead of raw embd_id.
  • Task cancellation is now authorized against dataset access.
  • The SDK preserves agent answer references.

Upgrade Notes

Note the version anomaly: v0.27.3 was published after v1.0.0-rc2 but carries a lower semver, so confirm which release line you are tracking before upgrading. Provision Apache Kvrocks 2.16.0 and run the database migration for the new chat message/reference tables before starting services, since the Go cache layer fails fast when Kvrocks is unavailable. Audit any custom parsers, chunkers, or scripts that relied on the old parser output format or the removed CLI insert commands.

The Two Releases

v1.0.0-rc2

v1.0.0-rc2

Oct 16, 2026· 9 days ago

This release is a Release Candidate 2 for version 1.0.0. The provided release notes are minimal and do not contain detailed changelog entries, feature lists, or specific updates.

Full release notes
v0.27.3

v0.27.3

Oct 23, 2026· 2 days ago

v0.27.3 is a massive release centered on porting the Agentic RAG stack to Go, alongside a broad ingestion/parsing refactor and a cache backend migration from Valkey 8 to Apache Kvrocks 2.16.0. It adds several new data-source connectors (Feishu Wiki, Zotero) and model providers (AnonRouter, API-Route, DaoXE, Cheaper Inference, GPT-5.6, recent Gemini models), while shipping a large volume of parser, retrieval, agent, security, and UI fixes. Numerous changes harden authorization/SSRF handling and improve DeepDoc/ONNX inference performance and memory behavior.

Full release notes