Skip to main content
Released December 2, 2025 · Alpha · View on GitHub This alpha line was finalized as 0.1.0. The last alpha before the beta is the largest of the point releases: 72 merged PRs. Two major connectors — Confluence and SharePoint — graduate to full sync support, Azure-hosted AI models (OpenAI, Claude, Grok, DeepSeek) join the provider list, and the indexing pipeline gets a substantial overhaul: reordered stages, batched embedding writes, faster Qdrant operations, and more file formats converted through the unified blocks schema. The rest is a wide stabilization pass across query, citations, viewers, and the graph layer.

​
Highlights

​
Confluence connector

Confluence spaces and pages can now be synced and searched as a first-class connector (#1017 — @shrikantNemiwal1).

​
SharePoint connector

SharePoint sync arrives (#1018 — @R0gue-one), with a display-name typo fix (#980 — @vishwjeet-pawar03).

​
Azure AI models and OpenAI-compatible embeddings

Azure-hosted models — OpenAI, Claude, Grok, and DeepSeek — are now supported as LLM providers (#1048 — @abhishekgml1991), and any OpenAI-compatible embedding endpoint can be configured (#1033 — @vishwjeet-pawar03). Model configuration also gained explicit context-length settings, used by the query pipeline instead of hard-coded limits (#1008, #1059), and a max_tokens attribute for Anthropic models (#1027 — @kaushal1999).

​
Indexing pipeline overhaul

The indexing pipeline was reordered (#975 — @kaushal1999), document additions are now batched (#982), create_embeddings was refactored (#992), and Qdrant operations were sped up (#996 — @abhishekgml1991). Blocks conversion expanded to more formats: PyMuPDF and Azure Document Intelligence output (#981), PPT (#1046 — @jain-ayush-wyld, first contribution), TXT parsed directly as Markdown (#990), and SVG images in HTML/Markdown (#954). Deadlock handling in indexing and fixes for empty-record indexing round it out (#1045 — @kaushal1999), plus a timeout on Docling PDF processing (#991) and a fix for re-indexing the same document (#930).

​
Platform settings and refreshed branding

A new platform settings page (#977 — @vishwjeet-pawar03), image previews for records (#1000), a new logo (#1006), and a replaced sign-in page image (#1031).

​
Connectors

  • Google Workspace — fixed multi-user sync for the Google and Drive connectors (#1028 — @vishwjeet-pawar03).
  • OneDrive — sync patch (#984 — @R0gue-one).
  • ServiceNow — fixed permissions, role hierarchy, and attachment streaming (#993 — @shrikantNemiwal1).
  • Outlook — fixed record streaming (#1022 — @shrikantNemiwal1).
  • BookStack — role handling patch, documentation fix, and a fix for records not visible to new users (#966, #983, #1069 — @R0gue-one).
  • Re-index now correctly emits and processes failed events (#1030 — @shrikantNemiwal1), and the reindex button is visible for failed records (#1042 — @vishwjeet-pawar03).

​
Agents & AI

  • Query responses that come back malformed now trigger reflection, with a fallback to reflection when JSON parsing fails (#1020, #1060 — @kaushal1999).
  • Retrieval limit reduced to 400 chunks after a tool call (#1023 — @kaushal1999) and a confidence-score bug fixed (#1038 — @vishwjeet-pawar03).
  • Citation improvements: Chinese square-bracket citations handled (#1005 — @abhishekgml1991), Excel citation highlighting fixed (#1035), and web-URL citation text matching added and then enhanced (#1036, #1039 — @vishwjeet-pawar03).
  • Fixed answer regeneration (#1047), streaming (#1007 — @kaushal1999), and metadata-enhancement bugs in search-result flattening and domain extraction (#1019, #972).
  • The model name is shown with each response (#1051 — @vishwjeet-pawar03), LLM health checks use a better test image and clearer messages (#1011, #1049 — @kaushal1999), and queries with no accessible records return a proper message (#1050).

​
Fixes

  • Image handling: images without URIs (#1016), 403 errors when fetching images in HTML/Markdown (#1024), and a cairosvg package bug (#998 — @kaushal1999).
  • Parser bugs: PyMuPDF processing (#1052), an uninitialized first_child_block_index variable (#1004), a wrong error type (#999), a wrong log in query (#1064), and one general bug fix (#1012 — @kaushal1999).
  • Graph layer: storage doc IDs are deleted from the graph DB on record delete (#1029 — @kaushal1999); permissions/permissionsToKb edge cleanup with migration code (#1044 — @stushar0711, first contribution); redundant code removed from the get_accessible_records query (#1054); duplicate code in base Arango that failed indexing (#1068 — @shrikantNemiwal1); Arango collection initialization moved out of the connector service (#1057 — @stushar0711).
  • A TypeError in etcd create_key (#1055 — @stushar0711).
  • Frontend: navbar flickering (#1041), PDF highlight brackets and chat message overflow (#1056), Google Docs/Sheets viewer (#1061), and file preview for filenames with special characters (#1066 — @vishwjeet-pawar03).
  • Uploads and routing: URL normalization in the backend (#970), files with the same name but different MIME types allowed (#1043), and enhanced route validation (#1040 — @vishwjeet-pawar03).

​
Performance & reliability

  • Faster Qdrant operations (#996) and credential caching to avoid repeated lookups (#1026 — @abhishekgml1991).
  • Indexing deadlock handling (#1045) and a Docling processing timeout (#991).

​
Developer experience & infrastructure

  • Added pikepdf to pyproject (#979 — @kaushal1999) and fixed Dependabot-reported issues (#1073 — @abhishekgml1991).

​
New contributors

  • @stushar0711 made their first contribution in #1044
  • @jain-ayush-wyld made their first contribution in #1046
Full commit-level detail: GitHub release · Compare view