gordonmurray/firnflow

The cost efficiency of S3 with the speed of local RAM. A multi-tenant vector and full-text search engine featuring a tiered RAM -> NVMe -> S3 architecture for microsecond latency on top of object storage

What it solves

Firn is a multi-tenant vector and full-text search engine designed to eliminate the need for always-on, expensive search clusters. It allows teams to run high-performance search directly against cheap object storage (like AWS S3, Google Cloud Storage, or MinIO) while maintaining low latency through a tiered caching system.

How it works

Firn uses a tiered storage architecture to balance cost and speed:

  • L1 (RAM) and L2 (NVMe): A hybrid cache (via foyer) stores frequent query results for microsecond-scale retrieval.
  • L3 (Object Storage): The source of truth where data is stored in isolated namespaces using LanceDB.
  • Optional Object Cache: A local NVMe layer that caches the raw bytes LanceDB reads from object storage, speeding up novel queries that aren't exact repeats.
  • Indexing: It utilizes IVF_PQ indexes to make cold searches on object storage practical (reducing latency from ~25s to ~979ms) and BM25 indexes for full-text search.

Who it’s for

It is built for teams managing multi-tenant SaaS workloads, private RAG deployments, or those looking to replace heavy services like OpenSearch, Elasticsearch, or Vespa with a self-hosted, low-idle-cost alternative.

Highlights

  • Multi-tenant by Design: Each namespace is physically isolated in its own object-storage prefix.
  • Hybrid Search: Supports vector, BM25 full-text, and fused hybrid search.
  • Late-Interaction Search: Supports multivector shapes (e.g., ColBERT, ColPali) for complex compositional queries.
  • CAS Consistency: Uses conditional-write primitives to ensure data integrity across concurrent writers.
  • Operational Visibility: Includes native Prometheus metrics to track cache hits and storage costs.

Related

  • Project
    endee-io/endee

    A high-performance AI search and intelligence platform that provides dense, sparse, and hybrid vector retrieval for RAG pipelines and AI agent memory.

  • Project
    sopaco/cortex-mem

    A Rust-based long-term memory framework for AI agents that uses a three-tier hierarchical architecture and a virtual filesystem to provide persistent, context-aware interactions.

  • Project
    codelibs/fess

    An enterprise search server built on OpenSearch that crawls and indexes content from websites, file systems, and various data stores for unified full-text search.

  • Project
    SeekStorm/SeekStorm

    SeekStorm is an open‑source Rust search engine offering native lexical (BM25, faceting, typo‑tolerance) and vector (ANN, quantisation, hybrid) indexes. It can be used as an in‑process library or as a multi‑tenant REST server, with client SDKs for Rust, Python, TypeScript, C# and Java. Designed for sub‑millisecond latency on a single commodity machine, it supports real‑time incremental indexing, sharded billion‑scale indexes, and a web UI for ad‑hoc queries.

  • Project
    NateBJones-Projects/OB1

    An infrastructure layer that provides a single, shared persistent memory database with vector search for all your AI tools, eliminating fragmented data across different AI clients.