Symbol-aware retrieval
A symbol graph links definitions to their references, so "who calls this function?" is answered from the code's actual structure rather than from text that looks similar.
Every capability, in one place
The landing page shows the highlights. This is the complete list: code intelligence, the dashboard, and the security model that keeps all of it on your own hardware.
Code intelligence
Retrieval that keeps up with your code
The retrieval engine tracks how your code connects, re-checks its answers as the code changes, and plugs into the AI tools you already use.
Citations that age honestly
If a cited file changed since it was indexed, the citation says so. The answer never silently points you at lines that moved.
Incremental indexing
Only changed files re-embed, so keeping a large repository current is fast and the nightly refresh stays cheap.
Answer quality, measured
Rate answers, and a built-in eval harness replays them after any model or index change and reports drift. This is the backbone of ongoing retrieval tuning.
Multi-repo Ask Pro
Ask one question across every indexed repository at once, with each citation tagged by repo. "How do the frontend and backend handle this?" comes back as a single answer. Standard from Pro up.
Git history answers
Commit messages are indexed alongside code, so "why was this changed?" and "when did this break?" are answered from the actual commits. Cloud indexers never see your history at all.
History at monorepo scale
Turn on full git history and every commit becomes searchable, with a dedicated engine option (Qdrant) that keeps history queries fast into the hundreds of thousands of commits. SourceVault downloads, verifies, and runs it for you: a second vector engine with zero operations work. See the measurement.
Module overviews
A generated repo map and per-module summaries answer "how does auth work overall" with an overview instead of fragments, written by your local model from your own code.
Cross-encoder reranking
A local reranker re-reads the top candidates against your actual question before answering, at sub-second cost. To be clear about what it buys you: the file-hit jump on our Express benchmark came from a separate retrieval fix, and the reranker's own effect was within noise on 30 questions. We keep it on by default and are re-measuring its benefit on a larger repository.
Works with your AI tools
An MCP server exposes the same engine to OpenClaw and any other MCP client, which get bounded, cited context instead of re-reading files. The server itself is local-only. Pair it with a local-model client and the whole loop runs offline; a cloud-backed client sends what it retrieves to its own vendor, by your choice.
The dashboard
A control plane your whole team can use
Everything ships with a browser dashboard. Connect your source control, manage repositories and models, and ask questions about your code without touching a terminal.
Connect your source control
Sign in to GitHub, GitLab, or Bitbucket once. Browse and autocomplete your repositories as you type, and clone private repos without per-clone credentials.
Webhook-driven reindex
Connect a GitHub, GitLab, or Bitbucket webhook and every push reindexes the repository automatically, so answers track the default branch without anyone remembering to sync. Standard from Pro up.
One-click indexing
Repositories index automatically on import. Update, sync, or switch branches per repo; a stale index is one click from fresh.
Search and Ask
Literal and semantic search with file-type filters, plus Ask mode for grounded answers where every citation clicks open to its source. History and archive are built in.
Built-in source viewer
Citations and search results open the full file in a syntax-highlighted viewer, with cited lines marked and scrolled into view. 15 languages, selectable light and dark code themes.
Local model manager
Pull, select, and uninstall Ollama models from the UI. The embedding model that powers search is protected from accidental removal.
Watches
Pin a question as a standing check. After every reindex it runs again and flags you if the cited answer has drifted. Useful for questions like "did the auth flow change this sprint?"
Runbook export
Pin good answers and export them as a markdown knowledge base generated from your own code, with every claim keeping its file-and-line citations.
Self-monitoring
Background polling keeps status, repositories, and models current without refresh buttons, and the health indicator flashes the moment anything needs attention.
Trust layer
Security-first by design
Nothing leaves the machine, and every control below is there to keep it that way.
100% local by design
Ollama generates the embeddings and ChromaDB stores the vectors, both on your own machine. The model that writes the answer runs there too.
Sentinel access policy
An enforcement gate sits between the index and every reader (dashboard, API, MCP). Files your policy excludes can't be read through SourceVault at all, no matter who asks or how.
Secrets redacted on the way out
A DLP pass scans everything leaving the engine, including answers, search previews, and file reads, and redacts keys, tokens, and credentials before delivery. Modern key formats are covered, and chunk boundaries are stitched so a secret can't slip through split in half.
AI-change provenance
Every commit is classified for AI authorship and security relevance, with signed attestations you can verify independently. When the index drifts, you can trace the drift to the commits that caused it and see whether a human or an AI wrote them.
Tamper-evident audit log
Every access decision lands in a hash-chained log with signed monthly anchors. Edits, truncation, and deletion of the log are all detectable, and logging fails closed rather than writing unprotected.
Locked-down dashboard
Authentication is enforced by default; there is no tokenless mode. A token is auto-provisioned at first boot, sessions get a one-click Lock, and a strict content-security policy plus loopback guard keep the control plane local.
Tokens instead of passwords
Access tokens are generated server-side and rotated from the UI in one click. Nobody types or chooses a credential, and rotation signs every other session out instantly.
Secrets never indexed
.env files, lockfiles, and dependency directories are excluded automatically so credentials never become searchable vectors, and Sentinel's DLP layer redacts anything that slips through on the way out.
No framework supply chain
The retrieval engine is built directly on auditable local services: ChromaDB for vectors, Ollama for models, and the optional Qdrant history engine, with no LangChain-style orchestration layer in between. That keeps the attack surface small and the code auditable.
HMAC-signed APIs
Search, file read, history search, and task endpoints require shared-secret signatures, with separate secrets so one leak does not expose the whole stack.
Repo confinement
Path-escape and symlink checks keep every read inside its repository, while file allowlists block binaries and unknown formats.
Deliberate deletion
Removing a repository requires typed confirmation and cleans up the working copy, vectors, metadata, and the indexed git history.