Manticore Search introduces Conversational Search, a layer built on top of full-text, vector, and hybrid search that lets a language model generate context-grounded answers, comparisons, and follow-up responses for e-commerce style shopping queries. The system works through CREATE CHAT MODEL and CALL CHAT SQL commands or a JSON HTTP API, using conversation history to build search queries, retrieving documents via vector search, and returning sources so answers can be traced back to actual products. Using the ConvApparel dataset (82,524 apparel products), a demo shop was built and benchmarked against 200 shopping queries, scoring 0.3650 Hit@3, 0.4250 Hit@5, 0.5250 Hit@10, and 0.2790 MRR — reported as the best among engines tested.
Table of contents
Word search, semantic search, and conversation are different tasksWhat we builtHow it worksThrough the HTTP APIWhat the system returnsHow retrieval worksWhere search ends and conversation beginsQualityConclusionQuestions this post answers
How do I set up conversational search with an LLM on top of Manticore Search?
Use the CREATE CHAT MODEL command to define a chat model with parameters like model, timeout, retrieval_limit, max_document_length, and a custom_prompt, then run CALL CHAT with the query, table name, model name, a conversation session ID, and the vector field name. The system retrieves relevant documents via vector search, feeds them to the language model as context, and returns an answer plus source IDs. See how daily.dev developers compare approaches for grounding LLM answers in real search results.
What retrieval quality did Manticore Search achieve on the ConvApparel shopping benchmark?
Manticore Search scored 0.3650 Hit@3, 0.4250 Hit@5, 0.5250 Hit@10, and 0.2790 MRR on a 200-query deterministic shopping benchmark built from the ConvApparel dataset, reported as the best result on each metric among the engines tested. The benchmark measures whether correct product IDs appear among retrieved sources rather than judging the wording of generated answers. Developers benchmarking search engines can track evaluation results like these on daily.dev.
Why doesn't full-text search alone answer comparison questions like which shoes are best for rain?
Full-text and hybrid search only return a ranked list of matching products; they don't interpret intent, compare options, or answer a follow-up question because that requires an additional conversational layer. A language model must sit on top of retrieval, using conversation history to build a complete search query and generating an explanation grounded in the retrieved product data. Track how search and RAG architectures evolve for handling conversational queries on daily.dev.