What infiniflow/ragflow shipped
Written by FoxPlug from the repository's own releases; not affiliated with InfiniFlow. Paste your own repository in the box below and FoxPlug writes the same weekly update for your product. This page is an automatic summary of the releases, pull requests and commits in github.com/infiniflow/ragflow; InfiniFlow did not write it and does not use or endorse FoxPlug. Every line links to the change it describes.
Get a weekly update like this for your product, free
Week of October 5, 2026
What shipped
- Parsers now emit structured TableData for every table instead of HTML cell strings, and chunkers read the structured rows directly, removing the legacy HTML re-parse path. Pull request #20741
- The first scheduled task now runs immediately on Start instead of waiting one full interval. Pull request #20714
- Retrieval now searches each tenant index once, stopping duplicate chunks when two or more datasets share an owner. Pull request #20731
- Chunk merge now counts its budget in runes rather than bytes, so CJK text no longer flushes early. Pull request #20690
- The Wikipedia tool now caps extracts by runes instead of bytes, so non-ASCII extracts are no longer cut inside a multi-byte rune. Pull request #19959
- TTS input is now capped at 500 Unicode code points instead of 500 bytes, and invalid UTF-8 plus paired markup tags are removed. Pull request #20344
- The Go image parser drops its hard edge and pixel limits and adds an opt-in raster byte ceiling, aligning it with the Python picture parser. Pull request #20745
- The Picture parser now runs the vision model when it is selected as the image
parse_method, instead of producing zero chunks. Pull request #20668 - A new upgrade guide covers the supported path from v0.20.x to v1.0.0-rc1, including a recommended route and a conservative route, and stresses starting, migrating and verifying each intermediate version. Pull request #20476
- GetWikiGraph now tokenizes its keyword to match pre-tokenized
content_ltks. Pull request #20722
Why it matters
Table handling moves from HTML strings to structured TableData, which changes how tables flow from parsers into chunkers. Retrieval, chunk merging, the image parser and several text caps are corrected, so duplicate chunks, early flushes and broken multibyte text are addressed. The new v0.20.x to v1.0.0-rc1 guide gives clear upgrade routes, including a conservative one for critical deployments.
Changelog entry
- Parsers emit structured TableData for tables and chunkers read the rows directly; the legacy HTML re-parse path is removed. Pull request #20741
- The first scheduled task runs immediately on Start instead of waiting one Interval. Pull request #20714
- Retrieval searches each tenant index once, preventing duplicate chunks across datasets with the same owner. Pull request #20731
- Chunk merge measures its budget in runes, not bytes. Pull request #20690
- The Wikipedia tool caps extracts by runes, not bytes. Pull request #19959
- TTS text is capped at 500 Unicode code points, with invalid UTF-8 and markup tags removed. Pull request #20344
- The Go image parser drops its hard edge and pixel limits and adds an opt-in raster byte ceiling. Pull request #20745
- The Picture parser runs the vision model when it is selected as the image
parse_method. Pull request #20668 - Added an upgrade guide from v0.20.x to v1.0.0-rc1 with recommended and conservative routes. Pull request #20476
- GetWikiGraph tokenizes its keyword to match pre-tokenized
content_ltks. Pull request #20722 - The dataset ingestion logs endpoint accepts type=sync and returns sync logs for connectors linked to that dataset. Pull request #20704
- The dataset ingestion logs endpoint rejects a
log_typeother than dataset or file. Pull request #19961
This week: tables now flow as structured TableData instead of HTML, and the legacy re-parse path is gone. Retrieval searches each tenant index once, so no more duplicate chunks. Scheduled tasks fire immediately on start. Chunk merge, Wikipedia extracts and TTS input now count run
A week of fixes and one structural change for users of the project. Parsers now emit structured TableData for every table instead of an HTML cell string, and chunkers read the rows directly, so the legacy HTML re-parse path is removed. Retrieval now searches each tenant index once, which stops duplicate chunks when multiple datasets share an owner. Scheduled tasks run immediately on Start rather than waiting a full interval. Several text limits now count runes rather than bytes: chunk merge, Wikipedia extracts and TTS input, which matters for CJK and other non-ASCII content. The Go image parser drops its hard edge and pixel limits in favour of an opt-in raster byte ceiling. The Picture parser runs the vision model when selected as the image parse_method. A new upgrade guide covers the path
Week of September 28, 2026
What shipped
- Agentic RAG ported to Go and exposed as a chat mode, with per-dialog failover support. Pull request #20503
- Grep and BM25 retrieval contracts re-exported for agentic tools with ~30x term matching speedup and document title indexing. Pull request #20542
- Chat pagination moved to database layer with COUNT and LIMIT/OFFSET for better performance on large datasets. Pull request #20450
- Fixed docx chunking regression where documents were collapsed into a single chunk. Pull request #20473
- DOCX XML character references now survive parsing after upgrading
office_oxideto v0.1.12. Pull request #20359 - Parallel workflow items now preserve state across pauses and resumes without rerunning completed steps. Pull request #20440
- Fixed Retrieval node's json output when search returns zero hits. Pull request #20394
- Picture OCR restored independently of vision enhancement setting for image parsing. Pull request #20424
- MinerU
server_urlnow forwarded through Go PDF parser for vlm-http-client and hybrid-http-client backends. Pull request #19445 - Go dependency downloader established as sole entrypoint, removing full-resource downloader. Pull request #20360
Why it matters
Agentic RAG moved to Go with performance improvements for retrieval, database-backed pagination scales better, and multiple parser regressions fixed to handle documents correctly. These changes improve execution speed, reliability, and stability for document processing and workflow execution.
Changelog entry
- Agentic RAG ported to Go and exposed as chat mode with per-dialog failover Pull request #20503
- Grep/BM25 retrieval contracts re-exported for agentic tools with ~30x term matching speedup Pull request #20542
- Chat pagination moved to database layer reducing memory usage Pull request #20450
- Fixed docx regression causing documents to collapse into single chunk Pull request #20473
- DOCX XML character references preserved through
office_oxideupgrade to v0.1.12 Pull request #20359 - Parallel workflow items preserve state across resume without rerunning Pull request #20440
- Retrieval node json output fixed for zero-hit search results Pull request #20394
- Picture OCR restored independent of vision enhancement setting Pull request #20424
- MinerU
server_urlforwarded through Go PDF parser backends Pull request #19445 - Refactored to use
type_kwdinstead ofknowledge_graph_kwdfor consistency Pull request #20475
Nightly 2026-10-04: agentic RAG now in Go with ~30x faster retrieval, database-backed pagination, and docx parsing fixes. Per-dialog failover and parallel state preservation improved workflow resilience.
This week's nightly brings agentic RAG to the Go backend with ~30x faster term matching and retrieval contracts re-exported for agent tools. Chat pagination now uses database-level operations for better performance at scale. We fixed critical parsing regressions in docx handling and XML character references, restored picture OCR independent of vision settings, and improved parallel workflow resilience with state preservation across pauses.
Week of September 21, 2026
What shipped
- Image chunk previews now stream directly to storage instead of batching in memory, reducing peak heap usage for large PDFs. Pull request #20165
- Image and table chunk cropping now runs in parallel across chunks instead of serially, improving chunker performance. Pull request #20183
- New model provider instances now auto-merge newly discovered models by default while preserving user-removed models. Pull request #20177
- Tool calling is now enabled by default for new provider model instances. Pull request #20195
- Re-parsing documents with existing chunks now shows a confirmation dialog before proceeding. Pull request #20210
- Ingestion log modal now polls for messages during parsing and stops at terminal states. Pull request #20206
- Shared chats and embedded bots can now access document images via beta tokens. Pull request #20185
- Rerank models now take effect during agent retrieval. Pull request #20197
- Disabled documents are now excluded from Wiki product datasets and graph entities. Pull request #20154
- Removed unused
tenant_llmtable from database. Pull request #20204
Why it matters
This week focused on performance improvements for document parsing with optimized image handling, better defaults for model provider configuration, and fixes for chat and retrieval features. Several refactoring changes cleaned up dead code in the agent execution path while fixes addressed dataset parsing, reranking, and document management issues.
Changelog entry
- Feat(chunker): stream cropped previews to storage instead of batching in memory Pull request #20165
- Perf(chunker): parallelize cropImageChunks across chunks Pull request #20183
- Feat(setting): auto-merge newly listed models into draft instance Pull request #20177
- Feat(setting): default tool calling on for new provider instances Pull request #20195
- Fix(dataset): confirm before re-parse Go documents with chunks Pull request #20210
- Fix(dataset): poll ingestion log while parsing and stop after terminal Pull request #20206
- Fix(router): accept beta tokens on document image routes for shared chats Pull request #20185
- Fix: rerank not working in agent retrieval Pull request #20197
- Fix: exclude disabled documents from Wiki products Pull request #20154
- Fix: remove table
tenant_llmPull request #20204 - Fix(setting): suppress PaddleOCR form autofill and default URL to first option Pull request #20189
- Fix: eval error Pull request #20158
Ragflow shipped parallel image cropping for faster chunking, streaming previews to reduce memory usage, auto-merge for model discovery, and fixes for reranking and document parsing.
Ragflow this week improved document parsing performance with parallel image chunk cropping and streaming preview uploads that reduce peak memory. Model provider setup now auto-merges newly discovered models while preserving removals, and tool calling defaults to enabled. Fixes address agent reranking, document re-parsing confirmations, ingestion log polling, and shared chat image access.