pinecone_search

Located at dapp/app/lib/mcp/tools/pinecone-search.ts, this tool turns user prompts into vector queries and returns a dual output: readable summaries for the LLM and JSON matches for the UI.

Runtime Traits

• Dynamic schema — The input schema is built from pineconeConfig, so enum values always match the configured indexes/namespaces.
• Embeddings — Uses Pinecone’s hosted multilingual-e5-large model for both query and corpus vectors, keeping deployments provider-agnostic.
• Namespace safety — pineconeConfig.isValidIndexNamespace guards against malformed requests before hitting the API.
• Result formatting — JSON includes matches with top scores and URLs. The text stream prefixes results with scores and emits gif-friendly <gif> tags for the UI renderer.

Pinecone Configuration

• dapp/app/config/pinecone.config.ts validates index names and namespace lists, and exposes helpers (getIndexNames, getNamespacesForIndex, isValidIndexNamespace) so tools can reject invalid queries before hitting the API.
• CLI utilities in dapp/pinecone (seed.ts, clear-namespace.ts, etc.) ingest JSON datasets such as gifs.json, rito-pics.json, and ritorhymes.json. They use Pinecone’s hosted embedding service (multilingual-e5-large) to stay provider-agnostic.
• MCP tools (pinecone-search.ts, agent-rap-verse/phases/gathering.ts) rely on the config to know which namespaces exist and to surface helper metadata (URLs, descriptions, meme categories).

Knowledge Hygiene

• Namespace awareness ensures the model never guesses index names; invalid combinations trigger descriptive errors before reaching Pinecone.
• Tool outputs mix text and JSON so transcripts remain lightweight while presenters and hydration hooks can extract structured information.
• JWT context (chain, token ID, address) is injected server-side for any tool that needs on-chain identity, reducing the risk of prompt spoofing.

// dapp/app/lib/mcp/tools/pinecone-search.ts
const pineconeSearchTool: Tool<PineconeSearchParams> = {
  name: 'pinecone_search',
  description: 'Perform semantic vector search in Pinecone indexes.',
  inputSchema: buildInputSchema(),
  async handler({ query, index, namespace = '__default__', topK = 5, includeMetadata = true, filter }) {
    const pc = getPineconeClient();
    const embeddingResponse = await pc.inference.embed('multilingual-e5-large', [query], { inputType: 'query', truncate: 'END' });
    const vector = embeddingResponse.data[0].values;
    const results = await pc.index(index).namespace(namespace).query({ vector, topK, includeMetadata, filter });
    const matches = Array.isArray(results.matches) ? results.matches : [];
    const text = textResult(buildSummary(index, namespace, matches));
    const json = jsonResult({ query, index, namespace, totalMatches: matches.length, topK, matches: normalize(matches) });
    return { content: [...text.content, ...json.content] };
  },
};

Presenter Notes

dapp/components/chatBot/ToolActivity/catalog/presenters/pinecone_search.presenter.ts adds context to the chips:

Usage Tips