Summary
Two related issues with the knowledge table MCP tools that make it impossible to filter rows or reliably assess table size from Claude Code.
Issue 1: relevance_list_knowledge_rows has no filter parameter
The relevance_list_knowledge_rows tool only accepts table_name, page, and page_size. There is no way to filter rows by field value via MCP.
Workaround: Use relevance_api_request with the flat /knowledge/list endpoint and a filters array in the body:
{
"method": "POST",
"endpoint": "/knowledge/list",
"body": {
"knowledge_set": "aws_contacts",
"filters": [{"filter_type": "exact_match", "field": "data.account_name", "condition_value": "Tapestry Inc."}],
"page_size": 20
}
}
Note: REST-style paths like /knowledge/sets/{name}/list, /knowledge/sets/{name}/documents/filter, /knowledge/sets/{name}/search, and /knowledge/sets/{name}/get_where all return 404 — so the flat /knowledge/list path is the only option.
Requested: Add a filters parameter to relevance_list_knowledge_rows so rows can be filtered by field value without dropping to raw API calls.
Issue 2: relevance_get_knowledge_table_info returns unreliable row_count
The row_count field returned by relevance_get_knowledge_table_info does not accurately reflect the number of documents in the table. In testing, a table with ~174 rows returned row_count: 1.
Impact: Claude Code cannot use row_count to decide how many pages to fetch, leading to either over-fetching or missed rows.
Requested: Fix row_count to return the actual document count, or document what it represents if it's intentionally counting something else (e.g. unique aliases).
Reproduction
- Table:
aws_contacts (~174 rows)
relevance_get_knowledge_table_info returned row_count: 1
relevance_list_knowledge_rows with page_size: 10 required 18 pages to retrieve all rows
- All filter/search REST-style paths on
relevance_api_request returned 404
Environment
- Claude Code with
relevance-ai-amazon MCP server
- Discovered during knowledge table scanning session (2026-04-04)
Summary
Two related issues with the knowledge table MCP tools that make it impossible to filter rows or reliably assess table size from Claude Code.
Issue 1:
relevance_list_knowledge_rowshas no filter parameterThe
relevance_list_knowledge_rowstool only acceptstable_name,page, andpage_size. There is no way to filter rows by field value via MCP.Workaround: Use
relevance_api_requestwith the flat/knowledge/listendpoint and afiltersarray in the body:{ "method": "POST", "endpoint": "/knowledge/list", "body": { "knowledge_set": "aws_contacts", "filters": [{"filter_type": "exact_match", "field": "data.account_name", "condition_value": "Tapestry Inc."}], "page_size": 20 } }Note: REST-style paths like
/knowledge/sets/{name}/list,/knowledge/sets/{name}/documents/filter,/knowledge/sets/{name}/search, and/knowledge/sets/{name}/get_whereall return 404 — so the flat/knowledge/listpath is the only option.Requested: Add a
filtersparameter torelevance_list_knowledge_rowsso rows can be filtered by field value without dropping to raw API calls.Issue 2:
relevance_get_knowledge_table_inforeturns unreliablerow_countThe
row_countfield returned byrelevance_get_knowledge_table_infodoes not accurately reflect the number of documents in the table. In testing, a table with ~174 rows returnedrow_count: 1.Impact: Claude Code cannot use
row_countto decide how many pages to fetch, leading to either over-fetching or missed rows.Requested: Fix
row_countto return the actual document count, or document what it represents if it's intentionally counting something else (e.g. unique aliases).Reproduction
aws_contacts(~174 rows)relevance_get_knowledge_table_inforeturnedrow_count: 1relevance_list_knowledge_rowswithpage_size: 10required 18 pages to retrieve all rowsrelevance_api_requestreturned 404Environment
relevance-ai-amazonMCP server