Crate sqlite_graphrag 

sqlite-graphrag

Local GraphRAG memory for LLMs in a single SQLite file — zero external services required.

sqlite-graphrag is a CLI-first library that persists memories, entities and typed relationships inside a single SQLite database. It combines FTS5 full-text search with sqlite-vec KNN over locally-generated embeddings to expose a hybrid retrieval ranker tailored for LLM agents.

CLI usage

Install and initialize once, then save and recall memories:

cargo install sqlite-graphrag
sqlite-graphrag init
sqlite-graphrag remember \
    --name onboarding-note \
    --type user \
    --description "first memory" \
    --body "hello graphrag"
sqlite-graphrag recall "graphrag" --k 5

Crate layout

The public modules group the CLI, the SQLite storage layer and the supporting primitives (embedder, chunking, graph, namespace detection, output, paths and pragmas). The CLI binary wires them together through the commands in commands.

Exit codes

Errors returned from errors::AppError map to deterministic exit codes suitable for orchestration by shell scripts and LLM agents. Consult the README for the full contract.

Modules

chunking

Token-aware chunking utilities for bodies that exceed the embedding window. Semantic chunking for embedding inputs (Markdown-aware, 512-token limit).

cli

clap definitions for the top-level sqlite-graphrag binary. CLI argument structs and command surface (clap-based).

commands

Subcommand handlers wired into the clap tree from cli.

constants

Compile-time constants: embedding dimensions, limits and thresholds. Compile-time constants shared across the crate.

daemon

Daemon IPC for persistent embedding model reuse across CLI invocations. IPC daemon: keeps the embedding model warm across CLI invocations.

embedder

Local embedding generation backed by fastembed. fastembed wrapper and per-process embedding cache.

errors

Library-wide error type and the mapping to process exit codes (see errors::AppError). Library-wide error type.

extraction

Hybrid entity extraction: regex pre-filter + candle BERT NER (graceful degradation). Entity and URL extraction pipeline (NER + regex prefilter).

graph

Graph traversal helpers over the entities and relationships tables. Entity graph traversal (BFS over memory_entities + relations).

i18n

Bilingual message layer for human-facing stderr progress (--lang en|pt, SQLITE_GRAPHRAG_LANG). Bilingual human-readable message layer.

lock

Counting semaphore via lock files to limit parallel invocations (see lock::acquire_cli_slot). Counting semaphore via lock files to limit parallel CLI invocations.

memory_guard

Memory guard: checks RAM availability before loading the ONNX model. Memory guard: checks RAM availability before loading the ONNX model.

migrations

namespace

Namespace resolution with precedence between flag, environment and markers. Namespace resolution layer (flag > env > “global” fallback).

output

Centralized stdout/stderr emitters for CLI output formatting. Single point of terminal I/O for the CLI (stdout JSON, stderr human).

parsers

Dual-format argument parser: accepts Unix epoch and RFC 3339. Input format parsers (Markdown, YAML, plain text, timestamp).

paths

Filesystem paths for the project-local database and app support directories. XDG/cwd path resolution and traversal-safe overrides.

pragmas

SQLite pragma helpers applied on every connection. SQLite PRAGMA helpers applied at connection open and on each transaction.

storage

Persistence layer: memories, entities, chunks and version history. SQLite persistence layer: sub-modules for each domain table group.

tokenizer

Real tokenizer of the embedding model for accurate token counting and chunking. Token-count utilities for embedding input sizing.

tz

Display time zone for *_iso fields (flag --tz, env SQLITE_GRAPHRAG_DISPLAY_TZ, fallback UTC). Display timezone for *_iso fields in JSON output.

Statics

SHUTDOWN

Signals that a shutdown signal (SIGINT / SIGTERM / SIGHUP) has been received.

Functions

shutdown_requested

Returns true if a shutdown signal has been received since the process started.