Open source · Built in Rust

The AI database
that reasons

ReasonDB introduces Hierarchical Reasoning Retrieval — LLM-guided tree traversal over your documents. Not chunks. Not embeddings. Structure.

Get started →View on GitHub

$ brew tap brainfish-ai/reasondb-tap && brew install reasondb

100%Benchmark accuracy

6.1sMedian latency

90%Context recall

RAG fails at scale

Flat chunks can't model document hierarchy. Vector similarity can't resolve cross-references. Black-box retrieval can't be audited. Standard RAG wasn't built for complex, regulated documents.

Challenge

Standard RAG

ReasonDB HRR

Document structure

Hierarchy, subsections, tables

Flat chunks lose hierarchy

Hierarchical tree nodes — full structure preserved

Cross-reference chains

§3.3 → §3.1 → Appendix

No cross-ref graph

cross_ref_node_ids built at ingestion

Domain vocabulary

'big C' vs 'critical illness'

No per-cohort vocab translation

Per-cohort domain_vocab · auto-extracted

Auditability

Regulated industries, compliance

Black-box, no audit trail

4-phase trace · logged · replayable

Determinism

Same query → same answer

Non-deterministic by default

Query cache · identical results

Formula handling

Financial calculations in docs

LLM hallucinates arithmetic

Formula nodes → Python tool-call

Query accuracy

Regulated document benchmarks

55–70%typical

90%+avg context recall

tree · policy_doc_47.pdf847 nodes

policy_doc_47.pdf847 nodes · 3 cross-refs
├── §1  General Provisions
│   ├── §1.1  Definitions
│   └── §1.2  Scope
├── §3  Coverage
│   ├── §3.1  Base Benefitscross_ref target
│   └── §3.8  Waiting Periodscross_ref target
├── §4  Eligibility
│   ├── §4.1  Requirements
│   └── §4.2  Conditions● beam
│         confidence: 0.94 · uses §3.1, §3.8
└── Appendix A

Architecture

Structure, not chunks.

Standard RAG flattens documents into chunks ranked by embedding similarity. HRR preserves the original hierarchy as a tree — the LLM navigates it deterministically, with full cross-reference awareness built at ingestion.

Sections stored as tree nodes — parent/child relationships intact
Cross-references resolved at ingestion: §4.2 knows it links to §3.8
LLM traverses the tree via beam search instead of ranking flat chunks
Identical structure every run — no embedding drift, no approximation

trace · query_id: 7f3a9c6.1s total

Phase 1 · BM25 candidate selection0.2s

12 docs matched

0 LLM calls

Phase 2 · tree_grep structural filter0.8s

4 docs · 18 nodes shortlisted

0 LLM calls

Phase 3 · LLM ranking1.4s

top 2 docs selected by summary + snippets

1 LLM call

Phase 4 · Beam search3.7s

§4.2 + §3.8 (cross-ref) · confidence 0.94

3 LLM calls

Reasoning: "§4.2 defines eligibility. §3.8 cross-ref provides waiting period. Confidence: 0.94"

Auditability

Every decision.
Fully visible.

ReasonDB logs every phase of every query. See which nodes were selected, why, and with what confidence. Replay any query identically — critical for regulated industries.

Full 4-phase trace per query — stored and replayable
Export to OpenTelemetry, Splunk, or any SIEM
Query cache ensures deterministic replay
Designed for SOC 2, HIPAA, SEC, and FedRAMP audit requirements

“We’d spent months trying to solve policy retrieval accuracy internally. RAG wasn’t cutting it — only 60% accurate on insurance benchmarks. ReasonDB is exactly what we’d been trying to build.”
Enterprise Architect
Top 10 global insurer · name withheld

Quick start

Up in minutes

Single Rust binary. Deploy on your own infra. No data leaves your environment.

01 · Install

$ terminal

# Install via Homebrew
brew tap brainfish-ai/reasondb-tap
brew install reasondb

# Or run via Docker
docker run --rm --pull always -p 4444:4444 \
  -e REASONDB_LLM_PROVIDER=openai \
  -e REASONDB_LLM_API_KEY=sk-... \
  brainfishai/reasondb:latest serve

02 · Ingest & query

$ terminal

# Start the server
reasondb serve

# Create a table and ingest a document
curl -X POST localhost:4444/v1/tables \
  -d '{"name":"policies"}'

curl -X POST localhost:4444/v1/tables/policies/ingest/file \
  -F 'file=@policy.pdf'

# Query with full reasoning trace
curl -X POST localhost:4444/v1/tables/policies/query \
  -d '{
    "query": "SELECT answer, confidence, trace FROM docs REASON \'Total disability benefit conditions?\'"
  }'

Full documentation →View benchmarks

Ready to go beyond RAG?

Open source. Single binary. Deploy anywhere. Bring your own LLM.

Get started for free →Star on GitHub