File Search

A composite tool

Most Vectoralix tools register as a single MCP tool. File Search is the exception. One File Search row in your dashboard fans out into up to five independently named MCP tools at registration time — one per enabled capability. The AI client sees them as separate callables and can choose which one to invoke.

search_tools row  ─┬─>  list_categories       (capability)
                   ├─>  list_groups            (capability)
                   ├─>  full_text_search       (capability)
                   ├─>  list_files             (capability)
                   └─>  search_related         (capability)

The five capabilities

• List Categories — return the categories defined on the active version of the server.
• List Groups — return the file groups within (optionally) a category.
• Full-Text Search — keyword search across the active version's content with snippets around the match.
• List Files — list content rows in (optionally) a group, with previews trimmed to a configurable size.
• Search Related — given a content row, return related rows by relation graph.

Per-capability settings

Every capability fieldset on the configure page exposes the same three top-level fields plus capability-specific knobs:

• Enabled — when off, the capability is dropped from the MCP registration. The tool can ship with only two of five capabilities active.
• Tool name — the name AI clients see for this capability. Make it readable to an LLM (e.g. search_docs, list_knowledge_categories).
• Description — a short hint to help the model decide when to call this capability.

Capability-specific knobs

Capability	Settings
List Categories	None
List Groups	None
Full-Text Search	max_results (default 10), snippet_length (default 200 words)
List Files	preview_mode (words/paragraphs), word_count (default 50), paragraph_count (default 1), max_results (default 20)
Search Related	max_results (default 10), snippet_length (default 200 words)

Version awareness

File Search always reads the server's active version. Switch the active version and the next AI client call returns results from the new snapshot — no redeploy of the tool itself is required, since tools themselves are not versioned. If you import new content but forget to cut a new version, the live content table updates while File Search keeps serving the older snapshot. See Versioning & Releases for the full distinction.

Typical AI-client call chain

AI clients are expected to chain capabilities. Categories give the broadest picture, groups narrow that down, and full-text search or list-files surfaces the actual content. A typical session looks like:

1. list_categories       ->  pick a relevant category
2. list_groups           ->  narrow to a group inside it
3. full_text_search      ->  find specific content
4. list_files / read     ->  get the full text the model needs

Safety and limits

• No external network calls and no sandbox concerns — File Search runs entirely against the tenant database.
• Defaults (10 results for full-text, 200-word snippets) keep context-window usage bounded; you can override per capability.
• Results are scoped to the server's active version only.
• Disabled capabilities never reach the MCP registration, so an AI client cannot call them by name.