Skip to main content
Conversational search is an AI-powered search feature that allows users to ask questions in everyday language and receive answers based on the information in Meilisearch’s indexes. Use conversational search when:
  • Users need easy-to-read answers to specific questions
  • You are handling informational-dense content, such as knowledge bases
  • Natural language interaction improves user experience
Use traditional search when:
  • Users need to browse multiple options, such as an ecommerce website
  • Approximate answers are not acceptable
  • Your users need very quick responses
Conversational search is still in early development. Conversational agents may occasionally hallucinate inaccurate and misleading information, so it is important to closely monitor it in production environments.

​
Conversational search user workflow

​
Traditional search workflow

  1. User enters keywords
  2. Meilisearch returns matching documents
  3. User reviews results to find answers

​
Conversational search workflow

  1. User asks a question in natural language
  2. Meilisearch retrieves relevant documents
  3. AI generates a direct answer based on those documents

​
Implementation strategies

​
Retrieval Augmented Generation (RAG)

In the majority of cases, you should use the /chats route to build a Retrieval Augmented Generation (RAG) pipeline. RAGs excel when working with unstructured data and emphasise high-quality responses. Meilisearch’s chat completions API consolidates RAG creation into a single process:
  1. Query understanding: automatically transforms questions into search parameters
  2. Hybrid retrieval: combines keyword and semantic search for better relevancy
  3. Answer generation: uses your chosen LLM to generate responses
  4. Context management: maintains conversation history by constantly pushing the full conversation to the dedicated tool
Follow the chat completions tutorial for information on how to implement a RAG with Meilisearch.

​
Model Context Protocol (MCP)

An alternative method is using a Model Context Protocol (MCP) server. MCPs are designed for broader uses that go beyond answering questions, but can be useful in contexts where having up-to-date data is more important than comprehensive answers. Follow the dedicated MCP guide if you want to implement it in your application.
⌘I