Complete Guide to Wiki Tools in TencentDB Agent Memory

The TencentDB Agent Memory repository exposes seven read-only Wiki tools—get_info, search, list_pages, read_page, get_graph, list_raw, and read_raw—that LLM agents invoke via HTTP endpoints to query wiki metadata, content, and knowledge graphs without write permissions.

TencentDB Agent Memory provides an LLM-agnostic interface for interacting with documentation wikis through the Knowledge Tools API. These tools are defined in MemoryKnowledge/src/routes/tools.ts and enable secure, read-only access to wiki resources, allowing agents to search documentation, traverse knowledge graphs, and retrieve source files programmatically.

Tool Architecture and Endpoints

Discovery and Execution Interface

All Wiki tools are exposed through two primary HTTP endpoints implemented in MemoryKnowledge/src/routes/tools.ts. The POST /tools/list endpoint returns the available tool catalog for a specific wiki resource, while POST /tools/call handles the actual execution. Every request requires an x-tdai-service-id header for service authentication and a knowledge_id parameter to identify the target wiki resource.

Whitelist Security Model

The system enforces a strict whitelist at line 70 of the tools route file. The WIKI_TOOLS registry contains an array of HttpToolDef objects that explicitly define permitted operations. Any attempt to invoke an undefined tool results in a 403 error, ensuring agents operate within a secure, read-only boundary.

The Seven Available Wiki Tools

get_info

Retrieves comprehensive metadata about the wiki, including its name, processing status, and total page count. According to the source code in MemoryKnowledge/src/routes/tools.ts (lines 51-54), this tool maps to wikiService.get and returns core properties without accessing document content.

search

Performs BM25 full-text search across all wiki pages. This tool accepts a query string and optional limit parameter (default 20), returning the most relevant documents with relevance-ranked snippets. The implementation delegates to wikiMgr.search in the wiki engine (MemoryKnowledge/src/engines/wiki/index.ts).

list_pages

Returns a complete inventory of page references, including page IDs, titles, and file paths. This tool calls wikiService.pageLs to enumerate all accessible pages in the wiki structure (defined at lines 64-67).

read_page

Retrieves full content from one or more pages identified by ID or path. The tool accepts a refs array parameter and executes wikiService.pageReadMany to fetch the actual markdown or text content (lines 69-74).

get_graph

Returns the knowledge-graph representation of the wiki, including nodes, edges, and community clusters. This tool invokes wikiMgr.graph to expose structural relationships between documents (lines 76-80).

list_raw

Enumerates the original source files uploaded to the wiki, such as markdown files and attachments. This maps to wikiService.rawLs and provides access to the underlying file inventory (lines 82-84).

read_raw

Reads the complete content of specified raw source files. Using wikiService.rawReadMany, this tool accepts file path references and returns the unprocessed source material (lines 86-90).

Practical Usage Examples

Listing Available Tools for a Wiki

To discover which tools a specific wiki supports, send a POST request to the discovery endpoint:

curl -X POST https://<your-host>/v3/knowledge/<team-id>/tools/list \
     -H "Content-Type: application/json" \
     -H "x-tdai-service-id: <service-id>" \
     -d '{"knowledge_id":"my-wiki-id"}'

Response excerpt:

{
  "knowledge_id": "my-wiki-id",
  "type": "wiki",
  "name": "Project Design Docs",
  "status": "ready",
  "tools": [
    {
      "name": "get_info",
      "description": "获取 wiki 元信息（名称、状态、页面数等）。",
      "params": {}
    },
    {
      "name": "search",
      "description": "BM25 全文搜索 wiki 页面内容。用关键词查找相关文档。",
      "params": {
        "query": {
          "type": "string",
          "required": true,
          "description": "搜索关键词"
        },
        "limit": {
          "type": "integer",
          "required": false,
          "default": 20,
          "description": "返回结果数上限"
        }
      }
    }
  ]
}

Executing a BM25 Search

Retrieve relevant documents using the search tool:

curl -X POST https://<your-host>/v3/knowledge/<team-id>/tools/call \
     -H "Content-Type: application/json" \
     -H "x-tdai-service-id: <service-id>" \
     -d '{
           "knowledge_id": "my-wiki-id",
           "tool_name": "search",
           "params": {
             "query": "authentication flow",
             "limit": 5
           }
         }'

Response excerpt:

{
  "results": [
    {"page_id": "12", "title": "Auth Overview", "snippet": "..."},
    {"page_id": "34", "title": "Login Process", "snippet": "..."}
  ],
  "count": 2
}

Reading Specific Page Content

Fetch full document content using the read_page tool:

curl -X POST https://<your-host>/v3/knowledge/<team-id>/tools/call \
     -H "Content-Type: application/json" \
     -H "x-tdai-service-id: <service-id>" \
     -d '{
           "knowledge_id": "my-wiki-id",
           "tool_name": "read_page",
           "params": {
             "refs": ["/design/authentication.md"]
           }
         }'

Response excerpt:

{
  "items": [
    {
      "ref": "/design/authentication.md",
      "content": "# Authentication\n\nThe system uses ..."

    }
  ]
}

Implementation Architecture

When a client calls POST /tools/call, the router validates the tool_name against the whitelist, then dispatches to executeWikiTool. This function contains a switch statement mapping each tool name to specific service implementations:

• get_info → wikiService.get
• search → wikiMgr.search
• list_pages → wikiService.pageLs
• read_page → wikiService.pageReadMany
• get_graph → wikiMgr.graph
• list_raw → wikiService.rawLs
• read_raw → wikiService.rawReadMany

The WikiService and WikiSourceManager are injected via the ToolsRouteDeps interface, keeping the routing layer focused on HTTP handling while delegating business logic to MemoryKnowledge/src/store/wiki-service.ts and the search/graph engines in MemoryKnowledge/src/engines/wiki/index.ts. The isWikiId helper function in MemoryKnowledge/src/store/ids.ts determines whether a given knowledge_id qualifies for wiki tool access.

Summary

• Seven read-only tools are available: get_info, search, list_pages, read_page, get_graph, list_raw, and read_raw
• Tools are exposed via POST /tools/list and POST /tools/call endpoints in MemoryKnowledge/src/routes/tools.ts
• Strict whitelist enforcement at line 70 prevents unauthorized tool invocation with 403 errors
• Service layer maps tools to specific WikiService and WikiSourceManager methods
• Raw source files are accessible separately from processed page content through dedicated raw file tools

Frequently Asked Questions

What authentication is required to use Wiki tools?

All requests must include the x-tdai-service-id header and target a valid knowledge_id that passes the isWikiId check in MemoryKnowledge/src/store/ids.ts. The system validates these credentials before exposing the tool registry or executing any tool calls.

Can LLM agents modify wiki content using these tools?

No. The WIKI_TOOLS registry explicitly defines read-only operations. There are no write, update, or delete tools exposed through the Knowledge Tools API, and any attempt to call an undefined tool results in a 403 Forbidden error.

How does the search tool rank results?

The search tool implements BM25 ranking through wikiMgr.search in MemoryKnowledge/src/engines/wiki/index.ts. This algorithm scores documents based on term frequency, inverse document frequency, and field length normalization to return the most relevant wiki pages for a given query.

What is the difference between read_page and read_raw?

read_page accesses processed and indexed content through wikiService.pageReadMany, returning rendered documentation. read_raw retrieves the original uploaded source files (such as markdown or text files) via wikiService.rawReadMany, providing access to unprocessed source material without indexing overlays.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}