Vector Search. In-Database Reasoning. Production-Safe Agency. All beside your data.
FractalSQL transforms SQLite from a passive data store into an active agentic database. FractalSQL adds what traditional RAG (Retrieval-Augmented Generation) stops short of: reasoning over what it retrieves, and when you enable it acting on the result, whether that's running a generated query or executing a decision an agent computed, all inside the same database process.
By bringing reasoning and agency directly into the SQLite backend, FractalSQL enables Sovereign Data Intelligence: the ability to reason, plan, and act upon your data with the deployment topology under your control. Run fully on-prem or in your own containers with Ollama/vLLM for zero data egress, or point at your organization's cloud AI accounts (Bedrock, Azure OpenAI, Vertex), where your compliance posture requires it for managed-model scale. You control the trade, not the product.
| Traditional RAG Stack | The Sovereign Way (FractalSQL) |
|---|---|
| Mode Collapse: top-K search returns near-duplicates, starving the LLM of diverse context. | Scout Discovery: MMR-style diverse semantic search that discovers the data's real structure. |
| Fragmented Logic: app pulls rows, calls LLM, handles retries, and glues answers in middleware. | In-Database Reasoning: reasoning and embedding happen inside the SQLite host process itself. |
| Passive Retrieval: you ask a question, the DB returns rows, and you hope the LLM is correct. | Autonomous Agency: self-correcting SQL, loop detection, and trajectory forecasting. |
FractalSQL's docs follow a single linear path. Each step answers one question and hands off to the next. You don't need to read everything; follow the path.
- What is this and why do I care?: you are here. Sovereign Data Intelligence, in one page.
- How do I get the extension running in 60 seconds? → Getting Started (
.loadthe extension, then your first Scout search). - How do I apply this to my industry? → Starter Kits: runnable, industry-specific SQL scripts (SOC, FinTech, MedTech, Fleet, Smart-Cities, …).
- How does a specific agent work and what are its inputs? → Agent Recipes: the built-in agents, each as a recipe.
- How do I build a proprietary agent that isn't in the box? → Composition Guide: the design patterns behind the recipes.
New here? Step 2 is a one-command demo. Step 3 drops you into a vertical that looks like your problem. Step 4 is the reference you'll keep coming back to.
Depending on your role, you'll want to start in different places:
- AI Engineer: You want to improve RAG quality and reasoning. → Start with docs/features.md and docs/reasoning-setup.md.
- DBA / Security Architect: You care about stability, safety, and access control. → See the Safety & Governance guide.
- Product Developer: You want to build agentic features quickly. → Run the Docker Demo, then pick a Starter Kit.
Four tiers of SQL-callable primitives, composable into agents with plain SQL and application code — no server-side language required.
- Discovery: diverse, mode-collapse-free retrieval:
fractal_search(Sniper),fractal_search_explore(Scout). - Cognition: in-process LLM integration:
fractal_reason(Bedrock, Azure OpenAI, Vertex, Ollama),fractal_embed,fractal_text_to_sql. - Agency: self-correcting routines: the search/sql agents, plan/trajectory/loop-detection helpers — see the Agent Recipes.
- Analytics: fractal/dimension primitives:
fractal_dimension_dfa,fractal_dimension_boxcount,fractal_optimize_portfolio, and more.
Plus SQLite-native plumbing: fractal_vector as a canonical BLOB type,
fractalsql_set()/fractalsql_get() per-connection configuration, the background
fractal_vectorizer embedding pipeline, and the tamper-evident
fractal_ledger_* audit surface. The full function-surface reference with
runnable examples lives in sql/fractalsql--1.0.sql.
Reasoning is opt-in and provider-pluggable: the fractalsql-reasoning-http
plugin speaks the OpenAI chat-completions wire format, so any OpenAI-compatible
endpoint works. Without a plugin configured, Discovery and Analytics are fully
functional and Cognition/Agency return clean precondition errors.
The fastest path is one command: Docker if you just want to try it, or the setup wizard if you have a real sqlite3 CLI install already. See Getting Started for the full walkthrough.
docker compose up -d # or: docker build -t fractalsql-sqlite .
# then .load the extension and run your first Scout search# Or, on a real install (Linux/macOS):
curl -fsSL https://github.com/FractalSQLabs/fractalsql-sqlite/releases/latest/download/easy_install.sh | bash
# Windows: scripts/windows/easy_install.ps1Native installers: .deb / .rpm for Linux amd64/arm64 (glibc and musl
channels), a Windows .msi (x64), and a self-contained tarball
for macOS (arm64/x86_64). easy_install.sh (Linux/macOS) and
easy_install.ps1 (Windows) wrap all of these behind one interactive
wizard; no telemetry, everything stays local.
Because configuration is per-connection (fractalsql_set — SQLite has no
GUCs to persist), the wizard writes a load_fractalsql.sql bootstrap
snippet you pass to every session: sqlite3 -init load_fractalsql.sql mydb.sqlite.
Everything above is Community edition and fully functional on its own. Discovery, Cognition, and Agency don't depend on anything in this section. For regulated environments that need to **prove, not just tamper-evident, HMAC-sealed decision record: see Enterprise Tier for the full mechanism, including what the ledger can and can't prove.
| SQLite | Linux | Windows | macOS |
|---|---|---|---|
| 3.25+ | ✓ | ✓ | ✓ |
Any SQLite host that permits sqlite3_load_extension — the sqlite3 CLI,
Python's sqlite3 module (conn.enable_load_extension), better-sqlite3,
libSQL/Turso-compatible hosts, mobile SQLite bridges. Verified on AWS
Graviton, Apple Silicon, Ampere Altra, and Raspberry Pi.
License: Apache-2.0. See LICENSE. Third-party components are under
their own permissive licenses (BSD-2-Clause, MIT, and others) --
see THIRD-PARTY-NOTICES.md.
For enterprise editions, licensing, and support, contact enterprise@fractalsqlabs.com.
Follow the path above; the links below are the same steps, expanded.
- Getting Started: 60-second Docker / native install.
- Starter Kits: industry-specific runnable SQL scripts.
- Agent Recipes: the sixteen installable agents, each as a recipe.
- Composition Guide: build your own agent.
- Features: the full Capability Map and API reference.
- Reasoning Setup: LLM provider configuration (Ollama, OpenAI, Bedrock, Azure, Vertex).
- Text-to-SQL Setup: pipeline details and the security model.
- Vectorizer Setup: automatic embedding pipelines.
- Docker Demo: a one-command end-to-end demo.
- Agent Blueprint Gallery: the vertical demos and reference agents.
- Enterprise Tier: the tamper-evident decision ledger (CISO/audit).