You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Fabricated citations stripped; uncited answers withheld in strict mode
Model-failure fallback
Partial
Falls back to an extractive answer if the model fails before output; a mid-answer failure ends with an error
Multi-hop / graph retrieval
Planned
Evaluation harness and results
Done
See docs/RAG_EVALUATION.md; LLM-generated answers not credibly measured locally
Language models
Provider
State
Notes
Anthropic (Claude Sonnet 5.5, Opus 5.5)
Done
Live end-to-end evaluation (docs/RAG_EVALUATION.md); API version 2026-10-21, optional effort, org-level workspace header
Ollama (local)
Beta
Selectable provider; opt-in in the demo seed (--ollama). Not used for evaluation: small models on a CPU were slow and produced garbled output
OpenAI (GPT-5.4 mini, GPT-4.1 mini)
Done
Live end-to-end evaluation; max_completion_tokens on OpenAI's API, automatic retry without temperature for default-sampling-only models
Gemini, Azure OpenAI, vLLM / OpenAI-compatible, Hugging Face TGI, custom
Beta
Factory and request formats unit-tested; no live keys used. Azure still sends max_tokens
AWS Bedrock
Beta
Converse API still sends temperature; current Claude models may reject it there (unverified)
Keys encrypted at rest, never returned by the API
Done
Tested
Per-org provider allow-list, fallback chain
Done
Any proprietary or open-source model
Done
Hosted APIs, Azure/Bedrock, Ollama, vLLM, TGI and any OpenAI-compatible server; any model id can be typed; agents can override the model
Live model discovery
Done
Lists fetched from each provider's API (OpenAI-compatible /v1/models, Anthropic Models API, Gemini, Ollama), stored as suggestions and refreshed daily by the worker. Verified live against OpenAI and Anthropic; Gemini and Ollama parsing unit-tested. Azure and Bedrock take typed ids
Platform model fallback
Done
PLATFORM_LLM_FALLBACK with Anthropic and/or OpenAI keys for organisations without their own provider; shown in Settings → AI providers
Chat and collaboration
Capability
State
Notes
Streaming chat (SSE) with citations and source previews
Fetches of gzip-encoded pages fixed (they failed to decode and were skipped)
GitLab (gitlab.com or self-managed; token optional for public projects)
Done
Live check 2026-11-02 against public gitlab-org/project-templates/express without a token (project, recursive tree, raw blobs). Token and self-managed paths: stubbed tests only
GitHub, S3-compatible storage
Beta
OAuth 2.0 framework (authorization code + PKCE S256, signed single-use state, token refresh with persisted rotation)
Done
Google, Microsoft identity platform, Dropbox, Notion, Atlassian 3LO, Slack. Tested end to end against stubbed vendor endpoints; no vendor's real consent screen has been exercised
Google Drive, OneDrive, SharePoint, Dropbox, Microsoft Teams
Beta
OAuth only. Need the vendor's OAuth app (OAUTH_* env vars); shown as "needs setup" until configured. Implemented and unit-tested against stubbed APIs; not verified live
Notion, Confluence Cloud, Slack
Beta
OAuth, or a token without an OAuth app (Notion internal integration token; Confluence email + API token; Slack xoxb- bot token). Not verified live
Email (IMAP)
Beta
SSL/TLS or STARTTLS with username + (app) password; each message becomes a Markdown document. Tested against a fake IMAP server only. No IMAP OAuth (XOAUTH2), so Gmail and Microsoft 365 need app passwords
SaaS sync gaps
Partial
Slack thread replies are not fetched (parent messages show the reply count); Confluence Server/Data Center is not supported (Cloud API only); Dropbox team spaces outside the member folder are not listed; disconnecting deletes stored tokens but does not revoke the grant at the vendor
Operations and platform
Capability
State
Notes
Connections command center (live map, list, data flow, SSE)
Agents with tools that inherit the caller's permissions
Done
Malformed tool input from a model is reported back to it instead of failing the run; live run verified on Claude
Agent tool approvals
Done
An approved tool runs in the background with the requester's permissions and the agent finishes its answer; the requester is notified and the run view follows the approval live. Sources gathered before the pause are re-authorized on resume
Approvals lifecycle
Done
Requesters can withdraw pending requests; pending approvals expire when their connection is deleted; nobody can decide their own request
Analytics, global search, notifications
Done
Notifications currently cover failed background jobs only
API keys (scoped, revocable), signed webhooks with test delivery
Done
Audit log (append-only, redacted, CSV export)
Done
Billing readiness (plans, limits, usage)
Partial
No payment provider; self-hosted shows "billing disabled"
Backups / restore scripts
Beta
deploy/backup.sh, deploy/restore.sh; not exercised
Testing gaps
Browser end-to-end suite (Playwright, frontend/e2e, 11 scenarios incl. mobile) runs green locally; its CI job is written but has not run on GitHub yet. No frontend unit tests.
No load, soak or concurrency testing.
Datasets list is unpaginated; /members is capped at 500.