An open ecosystem for search, retrieval, AI infrastructure, inference, and language models.
Portable, disk-first, open-source technologies designed to make search engines, vector databases, AI runtimes, and machine learning infrastructure composable, efficient, and independent from heavyweight platforms.
We believe AI infrastructure should be as portable as files on disk.
Search indexes shouldn't require clusters. Vector indexes shouldn't require specialized databases. Language models shouldn't require loading hundreds of gigabytes into memory.
Flatseek is building an ecosystem where knowledge, retrieval, model storage, inference, training, and AI applications share the same philosophy:
Portable. Disk-first. Open.
Flatseek Ecosystem
┌──────────────────────────────────────────────────────────────────────────────┐
│ DATA ENGINES │
│ │
│ Flatseek Flatvec │
│ Keyword Search Semantic Search │
└──────────────────────────────────────────────────────────────────────────────┘
┌──────────────────────────────────────────────────────────────────────────────┐
│ MODEL INFRASTRUCTURE │
│ │
│ Flatweight Flatrun │
│ Storage-native AI Streaming Inference │
│ Storage Runtime │
└──────────────────────────────────────────────────────────────────────────────┘
┌──────────────────────────────────────────────────────────────────────────────┐
│ MODEL TRAINING │
│ │
│ Flatbuild Flattune │
│ LLM Training LLM Fine-Tuning │
└──────────────────────────────────────────────────────────────────────────────┘
┌──────────────────────────────────────────────────────────────────────────────┐
│ APPLICATIONS │
│ │
│ Flatask Flatlens │
│ RAG Runtime Dashboard & Analytics │
└──────────────────────────────────────────────────────────────────────────────┘
Developer tools such as Flatbench and Flatperf support the ecosystem but are intentionally kept separate from the core architecture.
Every project can be used independently.
Together, they form a complete open-source platform for building modern AI systems.
| Project | Description |
|---|---|
| Flatseek | Disk-first full-text search engine comparable to Elasticsearch, OpenSearch, and Apache Lucene, designed for portable indexes without always-on clusters. |
| Flatvec | Disk-first vector database comparable to FAISS, Qdrant, Milvus, and Chroma for semantic search, embeddings, hybrid retrieval, and RAG. |
| Flatweight | Storage-native model format comparable to GGUF and SafeTensors, designed for portable, streamable AI model storage and execution. |
| Flatrun | Streaming LLM inference runtime comparable to llama.cpp, vLLM, MLX, and Ollama, executing models layer-by-layer directly from storage instead of loading entire checkpoints into memory. |
| Flatbuild | Language model training framework comparable to NanoGPT, LitGPT, torchtune, and Hugging Face Transformers, providing reproducible tokenizer training, model training, checkpointing, evaluation, and SafeTensors export. |
| Flattune | Fine-tuning framework that prepares datasets from Flatseek and Flatvec, then automates supervised fine-tuning, evaluation, and model specialization. |
| Flatask | Retrieval-Augmented Generation (RAG) runtime comparable to LangChain, LlamaIndex, and Haystack, connecting search, retrieval, and LLMs into grounded AI applications. |
| Flatlens | Search, monitoring, and analytics dashboard comparable to Kibana, OpenSearch Dashboards, and Grafana for exploring knowledge and AI data pipelines. |
| Flatbench | Benchmark suite for evaluating search engines, vector databases, storage systems, and AI runtimes. |
| Flatperf | Low-level profiling and performance analysis toolkit comparable to perf, VTune, and Tracy, providing runtime instrumentation and optimization insights. |
Today's AI stacks are typically assembled from many independent systems.
Traditional AI Stack
Search Engine
+
Vector Database
+
RAG Framework
+
Model Runtime
+
Training Pipeline
+
Dashboard
+
Benchmark Tools
Multiple systems
Multiple formats
Multiple pipelines
Flatseek takes a different approach.
Flatseek Ecosystem
One Philosophy
Portable • Disk-first • Open • Composable
│
Search → Retrieval → Inference → Training → Analytics
│
Build only what you need.
Combine everything when you don't.
Each project solves a different problem while sharing the same architectural philosophy.
- Disk-first by design — Data and models are built to live on storage instead of permanently occupying memory.
- Portable formats — Move indexes and models across machines without rebuilding infrastructure.
- Composable architecture — Every project works independently while integrating naturally with the rest of the ecosystem.
- Open standards — No vendor lock-in. No proprietary ecosystem.
- AI-native workflows — Built for search, retrieval, inference, fine-tuning, and modern AI applications.
- Research-driven innovation — Exploring new approaches to search, storage, inference, and machine learning infrastructure.
- Build search engines that scale without heavyweight databases.
- Make semantic retrieval simple, portable, and efficient.
- Simplify production-grade RAG systems.
- Build storage-native AI infrastructure.
- Enable streaming inference for models larger than available RAM.
- Make model training and fine-tuning portable.
- Reduce infrastructure complexity through reusable formats.
- Build a unified ecosystem where knowledge, models, inference, and applications share the same foundation.
Every project is released under the Apache License 2.0.
Use a single project—or combine the entire ecosystem to build your own AI platform.
