How a request flows
through Infino.
At the core of Infino is an agentic search engine and agent runtime with identity, permissions, and audit. Built in Rust from the ground up.
Same boundary, every call.
Identity first, permissions enforced before data access, every step recorded.
Authenticate the caller
A request enters through the UI or API. The Authentication layer verifies identity and establishes the request context.
Route through the agent runtime
The Agents service holds session state, scratch memory, and the audit context that follows the request through the rest of the stack.
Authorize execution
Authorization enforces RBAC, scope, and policy against the plan before anything can reach storage or remote systems.
Resolve against data
Authorized requests resolve against Infino's storage stack — full-text, vector, and SQL on one copy of your data.
What each part does.
Every component shares one identity model, one audit trail, and one storage layer.
Authentication
Verifies identity for humans, agents, and services, then establishes the request context for everything downstream.
Agents
Agent runtime that holds session state, scratch memory, and the audit context attached to every downstream call.
Authorization
Applies RBAC, scope, and policy to the request plan, then dispatches to the storage layer.
CoreDB
Append-only, time-aware engine for raw records, embeddings, and context across vector, full-text, relational, and time-series.
Object Store
Cost-efficient persistence for raw data and frozen segments, decoupled from compute so retention and query scale independently.
Metadata Store
Tracks derived schema, type history, lineage, and segment metadata so every field change remains queryable.
Embedding & Reranker
Models running inside the engine for embedding and ranking without external API calls or version drift.
One copy
of your data.
A custom search engine that executes Elasticsearch queries on Parquet files so a single copy of your data runs full-text, vector, and SQL at object-storage scale.
Vectors live inside the Parquet files themselves. No Pinecone, no pgvector, no embedding sync job to babysit.
Full-text indexes are embedded in Parquet alongside the data. No Elasticsearch fleet to scale and re-index.
Nothing to mirror, replicate, or reconcile. The Parquet file is the index, the vector store, and the table.
S3, GCS, Azure Blob. Storage and compute scale independently — agents stay cheap as data grows.
Still Parquet. Other tools can read it. You aren't locked into a proprietary store to query your own data.
Identity, permissions, and audit apply once — to the one copy — instead of being re-implemented per system.