Changelog

What's new

• Group-based permissions — a rebuilt permission system gives admins fine-grained control over who can access what. Default groups keep every standard user covered. Group managers can manage membership without full admin access. Audit events are emitted for all group and membership changes.
• Craft artifact index and action receipts — Build sessions now maintain a searchable index of all generated artifacts. Every tool invocation records a structured receipt showing what was called, what arguments were used, and what it returned — visible in the session timeline.
• Per-model reasoning and temperature settings — individual model configurations carry their own defaults for reasoning effort, temperature, and context window. Admins pin different defaults per model. Users see the active model on the Retry button and can adjust settings from a session pane.
• SCIM 2.0 provisioning — enterprise identity providers manage users and groups automatically via the SCIM API. Includes Entra soft-delete handling, rename collision healing, and username-based matching for IdPs that use non-email usernames.
• SSO domain routing — login attempts route to the correct SSO provider based on verified email domain. Multiple providers coexist; each maps to one or more DNS-verified domains.
• Salesforce OAuth connector — connect to Salesforce with full OAuth credential lifecycle: authorization, token refresh, and permission sync via the standard connector setup flow.
• JWT auth settings in the admin panel — configure issuer and audience enforcement for JWT auth directly from the Security and Hardening page. Settings are DB-backed and take effect without a restart.
• Language picker — users select their preferred interface language from account settings. The preference persists across devices. CJK fonts and line-breaking are included.
• LLM gateway on Business plan — the OpenAI/Anthropic-compatible gateway is now available on Business tier deployments, not just Enterprise.
• Per-user default temperature and reasoning — users set their own defaults for temperature and reasoning effort. Choices apply automatically to new sessions and persist across devices.
• Invite-only toggle moved to Security and Hardening — the invite-only setting now lives alongside other access controls for a cleaner admin experience.

Improvements

• Notification severity tiers allow platform messages to be classified as informational, warning, or critical. The banner queue and bell icon adapt accordingly.
• Connector failure alerts are louder — admins see a prominent alert for connectors that are failing or have invalid credentials, not just a status badge.
• Connector credential capability reports record exactly which permissions are present or missing after each credential check, making misconfiguration diagnosis faster.
• SCIM login email is driven from the emails attribute, supporting IdPs that use a separate login email from the display address.
• GitBook connector now indexes page descriptions, handles rate limits gracefully, and parses all block types with a recursive renderer.
• Jira poll windows now work correctly in non-UTC timezones — previously, updates could be skipped when the host timezone offset caused poll windows to misalign.
• MCP OAuth authorization is now resumable — an interrupted authorization flow can be continued without starting over.
• Cache-write tokens are now priced and persisted in usage tracking alongside cache-read tokens.
• Readiness probe moved to /health/ready for cleaner separation from the liveness probe at /health.
• Unstructured package removed — PDF and document parsing is lighter and no longer depends on the heavy unstructured dependency.

Fixes

• Duplicate get_docs_sync_status endpoint definition removed — the conflict caused an ImportError that prevented the API server from starting on a clean checkout.
• OIDC token refresh self-deadlock resolved — concurrent requests no longer contend on the same session lock when refreshing an expiring token.
• Reasoning effort correctly sent as none for GPT-5.4+ tool calls over the chat completions API.
• MinIO pulled from quay.io with a pinned digest — eliminates intermittent pull failures from Docker Hub rate limits.
• Reasoning level chosen before the first message is now persisted correctly.
• Disabled chat preference actions no longer appear in the chat UI.
• MCP tools are stored once when a connect event fires two refreshes.

What's new

• LLM gateway — expose your centrally managed model credentials as an OpenAI-compatible or Anthropic-compatible API endpoint. Point Claude Code, Codex CLI, or any OpenAI-compatible tool at your Baudhik deployment and route all traffic through your configured providers. All gateway usage is metered and counts against token limits.
• Incognito chat — conversations in incognito mode are never written to durable storage. The session lives in cache with a one-hour TTL. Uploaded files are deleted with the session and never reach the search index. Admins enable it per workspace or per group.
• Usage and cost management rebuild — the admin usage dashboard now shows workspace totals, a searchable per-user spend table with a daily breakdown by model and flow, and group and per-user budget controls. NPR (रु.) is available as the display currency with a daily-refreshed exchange rate and a configurable margin, so you see costs in Nepali Rupees without manual conversion.
• S3-compatible storage connector — index documents from any S3-compatible store — MinIO, Ceph, Wasabi, or a custom endpoint — by supplying an endpoint URL in the connector configuration.
• Multi-prefix storage connector — one BaudhikKB or S3-compatible connector now indexes multiple folder prefixes, replacing the need for a separate connector per folder.
• Citation-mapped cross-references — documents can declare related documents in their metadata. When a cited document is retrieved, its declared cross-references are automatically included as context, improving answers that span multiple related provisions.
• Structured input widgets — assistants can present clickable option buttons or an inline form instead of asking structured questions as free text. The user's selection arrives as a normal conversation turn with no page navigation.
• PDF source preview — cited PDF sources open inline in the chat panel with the relevant passage highlighted.

Improvements

• Admin usage export now includes question count and session count per user alongside token costs.
• Error messages shown to non-admin users are now clean and actionable. Admins still see full diagnostic detail.
• API responses for document sets, tools, and MCP servers return only name and ID to non-admin users, keeping corpus topology private.
• MCP servers support custom connection headers. OAuth token refresh is single-flighted and Google-hosted servers now request offline access, so connections no longer die after an hour.
• Account continuity across email changes — an identity-provider email rename no longer strands users. Baudhik adopts the new address at login and prior addresses continue to match document access controls while connectors re-sync.
• Redis Sentinel now works for Celery beat and the result backend, enabling high-availability Redis deployments.
• Azure voice provider supports non-English speech with configurable recognition languages.

Fixes

• Keyword and query metadata fields now flow correctly into BM25 and embedding search. Previously excluded due to a key-name mismatch.
• Slack bot answers are grounded in document search again — a prior regression had let the bot answer without searching.
• Image-summarization and contextual-RAG LLM calls are time-capped, so a stalled provider stream can no longer wedge indexing workers.

What's new

• SSO-only login — password authentication can now be disabled entirely for SSO-only deployments. A toggle under Security settings removes the email/password form while SSO login and provisioning continue. Enabling it requires an active SSO provider and the last enabled provider cannot be disabled while password auth is off.
• Craft — AI agent mode — describe what you want in natural language and Craft builds it: documents, presentations, structured reports, or web applications. The agent works in an isolated sandbox with full access to your indexed knowledge. Sessions can be scheduled to run automatically with pre-approval controls and failure notifications.
• Expert Deep Research — the Research Persona now has an Expert mode. The orchestrator exhausts all planned research steps before stopping, and the final report uses a structured template with precise citations and a clear hierarchy of findings.
• Nepali form filling — the assistant can locate the right government form and fill it out from conversation context, returning a completed PDF ready for submission.
• SQL query tool — admins configure a read-only database connection and expose it as an assistant tool. The assistant queries the database directly and presents results as structured data.
• Box connector — indexes files, web links, and folder hierarchy from Box with permission sync via Client Credentials Grant and optional user impersonation.
• Azure Blob Storage file store — a first-class Azure Blob backend for deployments that don't use MinIO or S3. Authenticates via connection string, account key, or DefaultAzureCredential.
• Per-chat reasoning and temperature controls — the model selector now exposes a reasoning-effort slider, a temperature slider, and a context-window gauge. Choices persist per chat session.
• Billing plan token limits — per-user token rate limits set by the subscription plan are enforced independently of the global rate limiter.

Improvements

• Tool routing rewired: the assistant now always calls the legal dictionary first, escalates to judicial precedent search, then falls back to document search — in the correct order every time.
• Cross-encoder reranker added after fusion ranking so the specific legal provision a user asked about appears at the top of results, not buried behind topically-similar chunks.
• Query history now supports filtering by user, assistant, flow type, and message content.
• Re-indexing shows a live progress banner with per-connector failure details. Units that fail repeatedly auto-pause and can be resumed.
• Admin log export collects from every worker, runs asynchronously, and a page refresh mid-export re-attaches.
• Large spreadsheets stream row-by-row through a file-backed pipeline instead of being loaded into memory.

Fixes

• Retrieval relevance guards now correctly prefer the exact legal provision named in a query over topically-similar chunks from adjacent provisions.
• Devanagari search queries are always generated in Nepali script regardless of how the user typed their question.
• A failed prune no longer re-fires on the scheduler.
• SCIM-provisioned users can now complete their first SSO login. Entra ID soft-delete tombstones no longer corrupt user emails.

What's new

• Najir (judicial precedent) search — a dedicated tool that searches only the court-decision corpus. Invoked automatically after statutory search when case law interpretation is needed, keeping legislative and precedent results cleanly separated.
• Legal dictionary — the assistant looks up statutory definitions directly without searching the full document corpus, returning precise definitions with source citations.
• Full document citation preview — clicking a citation opens the complete source document in a modal, not just the retrieved passage. Scroll through the full text to see context above and below the cited section.
• Knowledge-scoped assistants — admins configure exactly which document sets, folders, and files each assistant searches. Different workspaces can be scoped to their own corpus.
• Document-type authority ranking — Acts and Regulations are ranked above Directives, Circulars, and Guidelines in search results. The governing law appears first.
• Folder hierarchy select-all — selecting a parent folder in the connector browser selects all nested folders and files in one click.
• Coding Agent — a sandboxed agent with file download and bash tool access, available as a standalone agent and as a tool the chat agent can invoke. Disabled by default; admins enable it from the admin panel.
• Durable chat runs — a chat or Deep Research run executes to completion on the server regardless of client disconnects. Reopen the chat from any tab and it replays what you missed, then continues live.
• Granular agent sharing — agents now have explicit owners, ownership can be transferred, and shares carry their own permission level.
• Scoped Personal Access Tokens — tokens can be restricted to specific API surfaces including read:chat, write:chat, and read:search.
• Multi-provider SSO from the admin UI — Google, OIDC, and SAML providers are created and managed from a new admin page without redeployment. Multiple providers can run simultaneously.
• Time-based search filtering — search can filter by document creation date and last-updated date.
• System announcements — admins broadcast a markdown announcement to all users from the Appearance settings. Banners queue in a single bottom-left stack.
• SIEM-exportable audit logging — normalized audit events covering authentication, admin configuration, access control, and impersonation. Compatible with Splunk, Sentinel, Elastic, and Chronicle.
• TLS across the data plane — Postgres, Redis, and OpenSearch connections now support server certificate verification and mutual TLS.

Improvements

• Contextual RAG (retrieval-augmented chunking) is now available regardless of deployment mode. Each chunk is enriched with surrounding document context before embedding.
• Response modes (Direct, Balanced, Expert) can be set per assistant as a default, in addition to per message.
• End-to-end indexing visibility — failed documents surface with their error, and a targeted reindex can be triggered per document without re-running the entire connector.
• Permission sync has dedicated admin tooling: view sync state, inspect what was synced, and capture exception traces.
• Conversation-aware search scoping — internal search decides which connected sources to cover per call based on the agent's routing instructions.
• Slack Enterprise Grid support, including cross-workspace indexing.
• Documents whose content hasn't changed since the last run are no longer re-indexed.
• Vespa fully removed. OpenSearch is now the only document index.

Fixes

• Citation preview shows complete document text, not only the retrieved chunk.
• Google Doc parsing is size-bounded, fixing OOM crash-loops on large documents.
• PDF text extraction switched to pypdfium2, eliminating heartbeat-timeout deaths on large PDFs.
• OIDC access tokens refresh on every authenticated request for any OIDC provider.

Everything in Baudhik AI

This entry documents the complete set of capabilities available from general availability.

Knowledge connectors

Connect your organisation's sources and keep them in sync automatically. Supported sources include Google Drive, Confluence, Notion, Jira, Slack, GitHub, web scraping, file upload, and the BaudhikKB document store. Each connector supports configurable sync frequency and credential management from the admin panel. Connectors that fail repeatedly auto-pause and notify admins.

Nepali legal search

Baudhik AI is built for Nepali legal and regulatory work. Every search query is automatically generated in Devanagari script to match the document index, regardless of how the user typed their question. The platform knows the current Nepal fiscal year, uses रु. notation for currency amounts, and applies the legal instrument hierarchy (Ain > Niyam > Nirdeshika > Paripatra) when ranking results and structuring answers.

AI chat with citations

Ask questions in natural language and receive answers with inline citations linked to the exact passage in the source document. Conversations are multi-turn with full context retention. Citations show the document title, section, and page — not a raw filename or storage path.

Deep Research

Activate Deep Research for complex questions that require synthesising information across multiple documents. The agent plans a research strategy, searches iteratively, and produces a structured report with full citations.

Assistants and response modes

Create assistant personas scoped to specific document sets, connectors, and system instructions. Choose between Direct (concise), Balanced (default), and Expert (detailed, structured, fully cited) response styles per message or as an assistant default.

Voice mode

Speak your questions and hear responses. Speech-to-text and text-to-speech are configured from the admin panel. Voice conversations are saved with playback intact.

Multi-model chat

Send a single prompt to multiple models in parallel and compare responses side by side. Pick the response you prefer to continue the conversation with that model.

Code execution

The assistant can write and execute Python code to solve analytical problems, perform calculations, and generate visualisations. Code runs in an isolated container with strict resource and timeout limits.

Projects and file upload

Organise work in projects. Each project has its own uploaded files and is scoped to its own context. Files are chunked, embedded, and made searchable alongside connector-sourced documents. Supported formats include PDF, Word, Excel, PowerPoint, plain text, and Markdown. Tabular files (XLSX, CSV) are indexed as structured content rather than flattened text.

Web search

Agents can search the internet for real-time information using configured web search providers. Web search results are presented alongside internal knowledge, with citations from both sources.

MCP tools

Agents can be extended with MCP-compatible tools and servers. OAuth and shared API key authentication are supported.

Admin controls

• User management — invite, deactivate, and assign roles. RBAC with group-based permission assignment.
• Connector monitoring — view sync status, error history, and trigger manual re-syncs.
• Embedding configuration — switch embedding models and manage the reindex process with impact estimates.
• Usage tracking — per-user LLM token consumption and cost reporting.
• Announcements — publish a site-wide banner to all users from the admin panel.
• Query history — search and review past conversations with filtering by user, assistant, and date.

Enterprise

• SSO via OAuth2 and SAML. SCIM 2.0 user and group provisioning.
• License management with staged expiry notifications.
• Audit logging and query history.
• Self-hosted on your infrastructure — documents and queries stay within your control.
• Hooks system for extending request lifecycle behavior at well-defined points.

Welcome to Baudhik AI

Baudhik AI is a knowledge platform for Nepali legal and regulatory work. Connect your documents and systems, ask questions in natural language, and get cited answers grounded in your own data.

Every answer traces back to the exact passage in the source document. Nothing is fabricated from general knowledge. Baudhik AI runs on your own infrastructure — your documents and queries never leave your environment.