Ground AI and agents in enterprise data

Agentic retrieval is retrieval that an AI agent directs. The agent decides when to search, which datasets to query, and how to refine a query after it reads the results. In Spice, agents call the built-in search and sql tools through the MCP server gateway or through a model with tools: auto.

RAG reduces hallucinations by grounding model responses in data retrieved at query time instead of relying only on training data. Models grounded in incomplete or outdated data are more likely to hallucinate, so retrieval quality and freshness matter as much as the model itself. Spice retrieves live, governed data with SQL and hybrid search so the context passed to the model stays accurate.

Hybrid search combines vector similarity search and BM25 full-text search. In SQL, call vector_search() and text_search(), then merge the two ranked lists with rrf() (Reciprocal Rank Fusion). Results are rows, so you can join and aggregate them like any other table. The /v1/search HTTP API runs vector search across one or more datasets.

No. Spice creates embeddings with local or hosted models and runs vector similarity search inside the runtime. A vector database usually needs other systems for text search, keyword search, and joins with structured data. Spice runs all of these in one runtime that you can deploy locally, in your cloud, or as a managed service.

Yes. Spice federates queries across databases, object stores, and APIs with standard SQL. One query can retrieve structured records and combine them with semantic search results. This grounds model responses in operational data without manual data engineering.

Use hosted models from providers such as OpenAI, Anthropic, and Amazon Bedrock, or run local models, through the Spice AI Gateway. Call a model in SQL with the ai() function, so retrieved rows go to the model in the same query. Use a prompt in ai() to classify, label, or summarize each search result. See LLM inference and AI model serving for how model serving works in Spice.

The rrf() SQL function merges the result sets with Reciprocal Rank Fusion. Each search can set a rank_weight, and time_column with recency_decay boosts recent rows. Give keyword search more weight for exact lookups, and vector search more weight for natural-language questions. To reorder the top results with a reranker model or a chat model, wrap the query in rerank().

Spice keeps a local copy of hot data with DuckDB, SQLite, or Spice Cayenne acceleration, so retrieval does not query each source. Store large vector sets in Amazon S3 Vectors, which holds millions or billions of embeddings without always-on compute. To answer repeated searches without running them again, turn on the search results cache.