Reference MCP Semantic Search

semantic_search

Search the Glintstone corpus across tablets, lemmas, signs, scholars, publications, named entities, and composites.

Arguments

Argument	Type	Required	Default	Description
q	string	yes	—	Free-text query: a phrase, lemma form, P-number, scholar name, etc.
types	list[string]	no	all	Restrict to these entity types: tablets, lemmas, signs, scholars, publications, entities, composites
limit	integer	no	5	Per-group result limit (max 100)
mode	string	no	hybrid	Retrieval mode: hybrid, lexical, or semantic

Retrieval mode

hybrid — Reciprocal Rank Fusion over both lexical and semantic result lists. RRF works by scoring each document by the sum of 1/(k + rank) across both lists, where k=60. This consistently outperforms either retrieval method alone, especially for domain terms that appear in both the controlled vocabulary (sign names, lemma forms) and natural language queries.

lexical — Full-text search over the indexed forms. Exact and near-exact matches. Use when you know the specific ATF form or P-number.

semantic — Vector similarity search over embedded representations. Use when you want conceptually related results rather than exact lexical matches.

Response envelope

semantic_search returns a ToolResponse[GroupedTablePayload]:

{
  "data": {
    "groups": [
      {
        "group_type": "tablets",
        "items": [{ "p_number": "P227657", "designation": "...", ... }],
        "view_all_url": "/artifacts?search=grain+distribution"
      }
    ]
  },
  "summary": "Found matching tablets in the Ur III administrative corpus...",
  "sources": [
    { "label": "CDLI", "annotation_run_id": 1 }
  ],
  "interaction_id": "42"
}

The cursor field in the response enables pagination — pass it as cursor in the next request to retrieve the next page. Cursors have a 5-minute TTL.

Example invocations

semantic_search(q="grain distribution Ur III")
semantic_search(q="lugal", types=["lemmas", "tablets"])
semantic_search(q="P227657")
semantic_search(q="Piotr Steinkeller", types=["scholars"])
semantic_search(q="Epic of Gilgamesh", types=["composites"], limit=3)

Attribution

Every result item carries source metadata. The sources array in the response lists the annotation runs that produced the data. Pass the interaction_id to submit_correction if any claim based on these results needs correction.