mongodb-search-and-ai

A guide for implementing, optimizing, and troubleshooting MongoDB search solutions across Atlas Search (lexical full-text), Vector Search (semantic), and Hybrid Search (combined), driven through the MongoDB MCP server. It is used to build search functionality for text queries such as autocomplete, fuzzy matching, and faceted search, for semantic similarity and RAG applications, for text containment and substring matching, and for filtering across many fields with variable combinations. Its core principles are to understand the use case before building, always inspect existing indexes and schema first, explain indexes and require explicit approval before creating them, optimize configuration for the specific use case, and support read-only mode by providing complete index JSON when create, update, or delete tools are unavailable.

The workflow runs in phases. Discovery uses list-databases, list-collections, collection-schema, collection-indexes, and atlas-inspect-cluster to understand data, fields, existing indexes, and the cluster's MongoDB version, asking clarifying questions when the request is vague. It then chooses a search type: Atlas Search for keyword matching with relevance scoring, fuzzy matching, autocomplete, faceted search, and language-specific analysis; Vector Search for semantic similarity, natural-language understanding, RAG, and cross-modal search; or Hybrid Search for combining approaches using $rankFusion (rank-based) or $scoreFusion (score-based) to merge pipelines.

For hybrid search it performs a version check, since $rankFusion requires MongoDB 8.0+ and $scoreFusion requires 8.2+, and does not proceed or consult the hybrid reference if the requirement is unmet. It then consults the appropriate reference files for lexical indexing and querying, vector search, or hybrid search before recommending indexes or queries. Execution explains the index in plain language, shows the JSON, asks for an index name and explicit approval, creates it via the MCP create-index tool (or hands over JSON in read-only mode), and runs queries with the aggregate tool, validating results. It explicitly warns against using $regex or $text for search use cases because they lack relevance scoring, fuzzy matching, and language-aware tokenization, and covers edge cases such as missing fields, missing indexes, and multiple candidate collections. Use it to map business search requirements to the right MongoDB search implementation.