Overhaul docs retrieval and web search quality

Replace the abandoned llms-txt-mcp/Chroma docs backend with an in-repo
MCP service: SQLite WAL + FTS5 + sentence-transformer embeddings,
transactional source replacement, persisted state across restarts,
singleflight refresh with conditional requests, hybrid lexical/semantic
ranking with exact-duplicate collapse, source/host filters, and
explicit-by-default content retrieval. Add docs_rebuild and a
docs-rebuild CLI command.

Add deterministic llms-full.txt snapshot generation for machine-local
menus with hash-validated provenance manifests; lifecycle commands
promote a local menu to its snapshot only when the manifest validates.
Switch public source profiles to content-bearing llms-full.txt feeds.

Improve web search: bounded provider fallback with per-attempt
diagnostics and cancellation, an optional Brave Search API provider,
strict SearXNG engine selection, capped link/media extraction, and a
real engine=browser renderer that routes every request through the
existing SSRF vetting while blocking WebSockets, non-GET traffic, and
private destinations.

Extend release checks with offline unit suites and isolated candidate
container tests for both images.
This commit is contained in:
2026-07-25 08:49:26 -07:00
parent 29bcb123fa
commit 51dceee224
60 changed files with 3207 additions and 107 deletions

View File

@@ -64,6 +64,9 @@ Only the variables below are part of the public configuration surface. Other
| `CONTEXT_KIT_WEB_SEARCH_PROVIDER` | `searxng` | Default `search_web` provider; fallback order depends on this provider |
| `CONTEXT_KIT_WEB_SEARCH_HTTP_TIMEOUT` | `15000` | HTTP timeout in milliseconds for search providers |
| `CONTEXT_KIT_WEB_SEARCH_MAX_RESULTS` | `10` | Default search result count when clients omit `limit` |
| `CONTEXT_KIT_WEB_SEARCH_MAX_PROVIDER_ATTEMPTS` | `4` | Maximum providers attempted for one search |
| `CONTEXT_KIT_WEB_SEARCH_PROVIDER_TIMEOUT` | `15000` | Per-provider diagnostic timeout in milliseconds |
| `CONTEXT_KIT_BRAVE_SEARCH_API_KEY` | unset | Optional Brave Search API fallback credential |
| `CONTEXT_KIT_WEB_SEARCH_CHROME_PATH` | `/usr/bin/chromium` | Chromium path inside the web-search image for Bing fallback |
| `CONTEXT_KIT_WEB_SEARCH_BROWSER_USER_AGENT` | bundled Chrome/Linux UA | User agent for the Chromium-backed Bing fallback |
| `CONTEXT_KIT_WEB_SEARCH_MCP_COMPAT_MODE` | unset | Set to `legacy` for MCP clients with weak tool-schema parsers |
@@ -74,7 +77,7 @@ Only the variables below are part of the public configuration surface. Other
| `CONTEXT_KIT_DOCS_SOURCES` | `config/sources.default.txt` | Space-separated source profile files |
| `CONTEXT_KIT_DOCS_MAX_GET_BYTES` | `75000` | Max bytes returned by docs retrieval |
| `CONTEXT_KIT_DOCS_EMBED_MODEL` | `BAAI/bge-small-en-v1.5` | SentenceTransformers embedding model |
| `CONTEXT_KIT_DOCS_PREINDEX` | `0` | Set to `1` to re-embed every source on container start |
| `CONTEXT_KIT_DOCS_PREINDEX` | `0` | Set to `1` to refresh stale/missing sources in the background on startup |
| `CONTEXT_KIT_DOCS_LOCAL_SOURCES_DIR` | `${CONTEXT_KIT_DATA_DIR}/local-sources` | Machine-local llms.txt tree mounted read-only into docs-mcp |
| `CONTEXT_KIT_DOCS_LOCAL_SOURCES_PORT` | `8769` | Loopback port inside docs-mcp for serving local source files |
@@ -136,6 +139,10 @@ same-ID `restart` does not apply a changed TTL; it takes effect only when a new
container is explicitly provisioned. When freshness matters for one task,
prefer `docs_refresh` instead of replacing the shared container.
Use `bin/context-kit docs-rebuild [SOURCE_URL ...]` after parser/model changes or
to force an atomic rebuild. Existing searchable generations remain available if
a source fetch, parse, or embedding step fails.
## Browser CORS
`context-docs` disables browser CORS by default. CLI assistants and server-side
@@ -172,3 +179,10 @@ For local llms.txt files, place content under
`http://127.0.0.1:8769/path/inside/local-sources/llms.txt` or another URL that
ends in `/llms.txt` or `/llms-full.txt`; that loopback URL is inside the docs-mcp
container, not exposed on the host.
Run `bin/context-kit docs-snapshot [--only DIRECTORY]` to materialize linked
local menus. Each successful directory gets `llms-full.txt` and
`llms-full.provenance.json`; cache validators live under
`${CONTEXT_KIT_DATA_DIR}/snapshot-cache`. `--offline` rebuilds only from that
cache. During `start`/`restart`, a local `/llms.txt` URL is automatically changed
to its sibling `/llms-full.txt` when that file exists.

View File

@@ -10,6 +10,11 @@ Context Kit is designed to be safe by default for local development.
- The web-search MCP image runs as the non-root `node` user.
- Web-search MCP sessions are stateless. Its HTTP front end permits only
loopback/internal Host values and rejects every supplied Origin with 403.
- Browser fetch intercepts each network GET, resolves it outside Chromium, and
blocks private/localhost addresses, non-GET requests, request-count overflow,
and byte-budget overflow. Redirect targets are checked independently.
- Search diagnostics contain bounded categorized error messages and never emit
the optional Brave credential.
- Repomix mounts only the current project read-only.
- Docs indexing stores data under `$HOME/.local/share/context-kit` unless you
override it.
@@ -27,6 +32,11 @@ Only index sources you trust enough to retrieve into an agent conversation. More
sources are not always better. Large or noisy docs can make retrieval slower and
less precise.
Docs source replacement is transactional. SQLite WAL state persists on the docs
volume, removed source profiles become inactive immediately, and full content is
not returned by default. Local snapshot provenance is stored separately from the
retrieval text so metadata does not pollute ranking.
## Code-Editing MCP Servers
Context Kit's default MCP servers either read remote content or mount the

View File

@@ -133,9 +133,15 @@ race result rendering and return no items even when Chromium sees Bing result
cards. The override waits for result cards and decodes current Bing redirect
URLs before handing results back to the upstream fallback registry.
`fetch_url` is different: in upstream `mcp-web-search` 1.3.0, `engine=browser` is
accepted but reserved for future support. It does not currently invoke Chromium;
URL fetching uses the HTTP extractor path.
`search_web` now returns bounded `diagnostics.attempts` entries. Check each
provider's `status`, `duration_ms`, `result_count`, and categorized error before
changing provider order. An optional Brave API fallback is enabled only when
`CONTEXT_KIT_BRAVE_SEARCH_API_KEY` is set.
`fetch_url engine=browser` invokes Chromium for JavaScript-rendered pages. Every
HTTP(S) GET is intercepted and fetched through vetted DNS addresses; non-GET
requests, private/localhost destinations, more than 100 requests, and more than
20 MiB total browser traffic are blocked. Use `engine=http` for ordinary pages.
## Docs Indexing Is Slow
@@ -147,11 +153,10 @@ Cloudflare and other large docs sets can take significantly longer than the
default source profile. Set `CONTEXT_KIT_DOCS_PREINDEX=1` only if you want
startup to eagerly embed every configured source.
## Docs Tools Say Index Manager Not Initialized
## Docs Sources Report Refresh Errors
If `docs_query` or `docs_refresh` returns `Index manager not initialized` while
`/status` still responds, the HTTP wrapper is up but `llms-txt-mcp` failed to
initialize its embedding model or Chroma database. Check the container logs:
If `docs_sources` reports `last_error`, the service keeps the previous generation
searchable and records the failed check. Check the container logs:
```sh
docker compose -p "${CONTEXT_KIT_COMPOSE_PROJECT:-context-kit}" -f compose.yml logs docs-mcp
@@ -173,7 +178,7 @@ sudo chown -R "$(id -u):$(id -g)" "$DATA_DIR/docs" "$DATA_DIR/models"
bin/context-kit restart
```
`bin/context-kit start` now pre-creates these directories and `doctor` reports
existing directories that are not writable by the current user. If an assistant
client reports `Session not found` after restarting `docs-mcp`, restart the
assistant so it opens a fresh Streamable HTTP MCP session.
`bin/context-kit start` pre-creates these directories and `doctor` reports
existing directories that are not writable by the current user. The docs MCP
uses stateless HTTP sessions, so clients do not retain a session ID across calls.
Use `bin/context-kit docs-rebuild` after fixing the underlying error.