Skip to main content
Agentic search is still an early, fast-moving space, and the agents built on top of it can hallucinate. LLMs may occasionally produce inaccurate or misleading answers even when the retrieved source documents are correct. Monitor responses closely in production, follow the hallucination reduction guide, and configure guardrails to minimize this risk.
Agentic search covers any natural-language search experience where a Large Language Model (LLM) decides when and how to query your data, then uses the results to answer a question. Meilisearch acts as the retrieval layer when implementing Retrieval Augmented Generation (RAG): a fast, relevant search that an agent can call as a tool. With proper configuration, such as system prompt engineering and guardrails, you can help keep responses based on your indexed data rather than the LLM’s general knowledge. This is similar to how Perplexity works: every answer comes with source documents so users can verify the information. Meilisearch brings the same pattern to your own data and your own agents.

​
Use cases

Agentic search covers a range of natural-language search experiences, from AI agents that call Meilisearch as a tool to conversational chat interfaces:

​
Agents with a search tool

Give any AI agent or LLM application the ability to query your Meilisearch indexes as part of a larger task, tool call, or multi-step reasoning process. The agent decides when to search, what to search for, and how to use the results. Example: A support agent looks up your product catalog before answering a compatibility question, then combines the results with its own reasoning.

​
RAG pipelines

Integrate Meilisearch as the retrieval layer in a broader RAG architecture that your agent or application controls. Meilisearch handles hybrid retrieval, while your code decides how to use the results and when to generate a response. Example: A product recommendation engine that retrieves matching products via Meilisearch, then uses a custom prompt to generate personalized suggestions.

​
Multi-turn chat

Build a full conversational interface where users ask follow-up questions and the agent maintains context across the conversation. This is ideal for knowledge bases, customer support, and documentation search. Example: A user asks “What models do you support?”, then follows up with “Which one is the fastest?” without restating the context.

​
One-shot answer summarization

Generate a single, concise answer to a user’s question without maintaining conversation history. This is useful when you want to display a summarized answer alongside traditional search results. Example: A user searches “How do I reset my password?” and gets a direct answer synthesized from your help articles, displayed above the regular search results.

​
How it works

Here is how a typical agentic search workflow looks like:
  1. Query understanding: The LLM transforms the user’s natural language question into optimized search parameters
  2. Retrieval: Search your indexes to find the most relevant documents using hybrid search for better relevancy
  3. Answer generation: The LLM generates a response using only the retrieved documents as context
  4. Source attribution: Every response can include references to the source documents used to generate the answer
Depending on your application needs, you may give more or less autonomy to the LLM at each given step. Stricter control makes for more predictable behavior (for example, translating a question into search parameters), while looser control allows the model to reason and find more creative solutions (for example, multi-hop reasoning).

​
Implementation strategies

@meilisearch/ai-sdk provides ready-made search tools for the Vercel AI SDK, so any agent or LLM application can call Meilisearch directly. This is the recommended way to build agentic and conversational search on top of Meilisearch for production applications. This approach provides search as native agent tools: you get static typing, in-process execution, and full control over each tool’s description, schema, and search parameters (filters, sorting, hybrid search). This approach allows for the most advanced retrieval configuration. Get started with the agentic search getting started guide.

​
Model Context Protocol (MCP)

Meilisearch has a built-in Model Context Protocol (MCP) server. It exposes four read-only search tools (listIndexes, describeIndex, searchInIndexes, and facetSearch) that an MCP client can call. MCP is the fastest way to get an agent talking to your data. It is a good fit when you want to experiment, or when you can add an MCP server to an agent but cannot change the agent’s code. This approach avoids writing tool code: the agent discovers your indexes and learns how to search them. The trade-off is control: the agent sees the full search API instead of a tool scoped to your use case, and tools are loaded dynamically, so they are not typed. To connect an MCP client, follow the MCP guide

​
Chats API (experimental)

The experimental /chats route consolidates retrieval, context management, and generation into a single, engine-managed endpoint. Use it only if you specifically need Meilisearch to handle the LLM call for you. See the Chats API guide for setup, chat patterns, streaming, and tools. Consult the chat completions API reference for the full list of supported parameters.

​
Choose an approach

Apps built with the Vercel AI SDK can also connect to tools provided by MCP. So you can start with MCP and switch to @meilisearch/ai-sdk tools when you need more control.