Crate docling_rag 

docling-rag: a pluggable Retrieval-Augmented-Generation subsystem built on the docling.rs document converter.

The pipeline is: source → convert to Markdown → chunk → embed → vector store → retrieve → (optionally) synthesize an answer. Every external dependency is a trait with swappable backends:

• Embedders (embed): Ollama (default), Gemini, local ONNX, or a deterministic hashing embedder for offline tests.
• Vector stores (store): SQLite (default), PostgreSQL + pgvector, or in-memory. Documents and chunks live in separate tables.
• Retrieval (retrieve): dense vector, sparse BM25, Hybrid (RRF), Multi-Query fusion, and HyDE.
• LLM (llm): OpenRouter (default model DeepSeek-V3).
• Sources (source): local folder (default), FTP, SFTP.
• Queues (queue): in-process, RabbitMQ, Redis pub/sub.

Configuration comes from the environment / a .env file via RagConfig.

use docling_rag::{RagConfig, Pipeline, RetrievalMode};

let cfg = RagConfig::from_env()?;
let pipeline = Pipeline::from_config(&cfg).await?;
pipeline.ingest_all().await?;
let hits = pipeline.query(RetrievalMode::Hybrid, "how does chunking work?", 5).await?;
for h in hits {
    println!("{:.3}  {}", h.score, h.chunk.text);
}

Re-exports

pub use config::RagConfig;

pub use error::RagError;

pub use error::Result;

pub use model::Chunk;

pub use model::Document;

pub use model::RetrievalMode;

pub use model::Scored;

pub use pipeline::Answer;

pub use pipeline::IngestOutcome;

pub use pipeline::IngestReport;

pub use pipeline::Pipeline;

pub use retrieve::Retriever;

Modules

api

REST API over a Pipeline: document info and search in every retrieval mode.

chunk

Markdown-aware chunking.

config

Configuration, loaded from the process environment (and a .env file).

embed

Pluggable embedding providers.

error

Error type for the RAG subsystem.

eval

Evaluation harness: sweep a matrix of {chunk_size, overlap, retrieval mode} over a labelled dataset and rank the configurations by retrieval quality.

llm

Pluggable chat/LLM provider, used for Multi-Query and HyDE retrieval and for final answer synthesis. The only shipped backend is openrouter.

math

Small, dependency-free vector maths used by embedders and stores.

metrics

Per-document processing metrics, captured during ingestion and stored in the document’s JSON metadata column under the "metrics" key — a plain JSON object, so new metrics can be added later without a schema migration.

model

Core data types shared across the pipeline: documents, chunks, scored hits.

pipeline

End-to-end orchestration: ingestion (source → convert → chunk → embed → store) and querying (retrieve → optional LLM answer synthesis).

queue

Pluggable message queue for decoupling document discovery from ingestion.

retrieve

Retrieval: dense vector, sparse BM25, and the advanced modes that combine or rewrite queries (Hybrid, Multi-Query fusion, HyDE).

source

Pluggable document sources. The default is a local folder; FTP and SFTP are available behind the remote-sources feature.

store

Pluggable vector store.