cascading-search.ts•3.15 kB
import {McpServer} from '@modelcontextprotocol/sdk/server/mcp.js';
import {Pinecone} from '@pinecone-database/pinecone';
import {z} from 'zod';
import {RERANK_MODEL_SCHEMA} from './common/rerank-model.js';
import {SEARCH_QUERY_SCHEMA} from './common/search-query.js';
const INSTRUCTIONS = `Search across multiple indexes for records that are
similar to the query text, deduplicate and rerank the results.`;
const INDEX_SCHEMA = z.object({
name: z.string().describe('An index to search.'),
namespace: z.string().describe('A namespace to search.'),
});
const RERANK_SCHEMA = z
.object({
model: RERANK_MODEL_SCHEMA,
topN: z
.number()
.optional()
.describe(
`The number of results to return after reranking. Must be less than or
equal to the value of "query.topK".`,
),
rankFields: z.array(z.string()).describe(
`The fields to rerank on. This should include the field name specified
in the index's "fieldMap". The "bge-reranker-v2-m3" and
"pinecone-rerank-v0" models support only a single rerank field.
"cohere-rerank-3.5" supports multiple rerank fields.`,
),
query: z
.string()
.optional()
.describe(
`An optional query to rerank documents against. If not specified, the
same query will be used for both the initial search and the reranking.`,
),
})
.describe(
`Specifies how the results should be reranked. Use a "query" with a "topK"
that returns more results than you need; then use "rerank" to select the
most relevant "topN" results.`,
);
export const SCHEMA = {
indexes: z
.array(INDEX_SCHEMA)
.describe('The indexes to search across. Records in each index should share a common schema.'),
query: SEARCH_QUERY_SCHEMA,
rerank: RERANK_SCHEMA,
};
export function addCascadingSearchTool(server: McpServer, pc: Pinecone) {
server.tool('cascading-search', INSTRUCTIONS, SCHEMA, async ({indexes, query, rerank}) => {
const initialResults = await Promise.all(
indexes.map(async (index) => {
const ns = pc.index(index.name).namespace(index.namespace || '');
const results = await ns.searchRecords({query});
return results;
}),
);
const deduplicatedResults: Record<string, Record<string, string>> = {};
for (const results of initialResults) {
for (const hit of results.result.hits) {
if (!deduplicatedResults[hit._id]) {
deduplicatedResults[hit._id] = hit.fields as Record<string, string>;
}
}
}
const deduplicatedResultsArray = Object.values(deduplicatedResults);
const rerankedResults =
deduplicatedResultsArray.length > 0
? await pc.inference.rerank(
rerank.model,
rerank.query || query.inputs.text,
deduplicatedResultsArray,
{
topN: rerank.topN || query.topK,
rankFields: rerank.rankFields,
},
)
: [];
return {
content: [
{
type: 'text',
text: JSON.stringify(rerankedResults, null, 2),
},
],
};
});
}