RAGatm Open
  • Open source · Apache-2.0
  • Local-first
  • Self-hosted
  • Cited answers

Your documents. Your AI. Your infrastructure.

An open-source RAG platform for analysing documents with locally hosted AI models, source-backed answers and control over your data.

Install the Compact profileView the code on GitHub

RAG against the machine

RAGatm Open answers your questions from your own documents. It retrieves the relevant passages, hands them to a language model running on your machine, and returns an answer that cites its sources.

Everything runs on your side: the application, the database, the models. In the qualified configuration, no document is sent to a remote AI provider, and the published code contains no fallback to an external service.

Three benefits

Answers you can verify

Every answer cites the document, the page or slide and the excerpt used. If a dependency is missing, the application says so instead of making things up.

Your data stays with you

Local models, internal networks without Internet egress, a single published port. You control hosting, backups and access.

Reasonable resources

The Compact profile is qualified on an 8-thread, 14 GB RAM machine with no GPU. Three open models, downloaded once.

The application

Real screenshots of the Open edition, on an isolated instance with fictitious documents.

Assistant answer to a question on the demo corpus, with source markers and the Sources panel listing the documents and excerpts used.
Locally generated answer, with its citations. (real generation on CPU, qwen2.5 1.5 B model, demo instance without reranker, ≈ 36 s; not a performance measurement)
Side panel of sources with an expanded excerpt showing the document, the page and the retrieved text.
Sources panel: document, page and exact excerpt. (real generation on CPU, qwen2.5 1.5 B model, demo instance without reranker, ≈ 36 s; not a performance measurement)
Document list (DOCX, PDF, TXT, Markdown) with “Ready” status, indexing column with the number of excerpts, banner showing the embedding pipeline and quotas.
The 8 demo corpus documents are ready, with the number of indexed excerpts.
“Drag and drop your files” zone listing PDF, Word, Excel, PowerPoint, CSV, HTML, Markdown, TXT and images, 50 MB per file limit.
Document upload: drop zone and formats announced by the server.
Dashboard with the sidebar (Home, Assistant, Documents, Business corpus, Spaces and members, administration) and the space selector.
Home after login: current space, navigation and shortcuts.
Assistant answer with citations displayed on a phone screen.
Cited answer on mobile. (real generation on CPU, qwen2.5 1.5 B model, demo instance without reranker, ≈ 36 s; not a performance measurement)

Screenshots taken on an isolated RAGatm Open instance (Open edition, Compact profile on CPU, reranker disabled for the capture), with the fictitious documents published in the repository and a demo account.

Four steps

The complete flow is described on the Architecture page.

  1. Import

    Drop your files into an isolated document space. Content is validated by its signature, extracted, then chunked.

  2. Index

    Each chunk gets a locally computed vector and joins PostgreSQL’s full-text index.

  3. Ask

    Hybrid retrieval (keywords and vectors), then reranking by a local cross-encoder.

  4. Read the cited answer

    A local model streams the answer, with typed sources for each statement.

Explore the architecture

Who is it for?

Organisations that want to query their documents without handing them to a remote service.

SMEs

Procedures, contracts, reports, technical documentation.

Law firms and regulated professions

Search and comparison across large case files, on controlled infrastructure.

Non-profits

Archives, bylaws, minutes, internal procedures.

Freelancers and researchers

Notes, publications, personal collections.

IT teams

Docker Compose deployment, API, contracts, modules.

See the use cases

Supported formats

The catalogue lists every format identified in the code, with a verified status for each.

Enabled by default
  • Plain text (.txt)
  • Markdown (.md)
  • PDF (text layer)
  • Word (.docx)
  • CSV
  • HTML
  • Excel (.xlsx)
  • PowerPoint (.pptx)
Configurable
  • Generic XML (.xml)
  • DXF drawing (.dxf)
  • MIME e-mail (.eml)
Optional
  • Scanned / image-only PDF (OCR)
  • JPEG image
  • PNG image
  • TIFF image
  • WEBP image
  • IFC model (.ifc)
  • DWG drawing (.dwg)

Statuses evaluated for the Compact profile as published. 44 formats listed.

Browse the format catalogue

Local AI and data control

What the published code allows, what is enabled by default, and what remains your responsibility.

Locally executed models

Generation, embeddings and reranking go through AI Gateway Open, which only calls local engines (Ollama, TEI).

No external provider

The only accepted value for the external fallback is "none"; startup checks reject any other configuration.

Partitioned network

Internal Docker networks, a single port published on loopback, no Internet egress in nominal operation.

Open, auditable code

Three Apache-2.0 repositories, hash-locked dependencies, base images pinned by digest.

Backups and data under your control

Documented backup and restore procedures; no model weight or data leaves your host.

Strong authentication

Mandatory TOTP second factor for every account, workspace roles, isolated spaces.

Controlled costs, developed in France

Free software with no per-user licence, running on a server without a GPU. Developed in France, with a European approach to technological control; no sovereignty certification is claimed.

Self-hosted software is not "compliant" or "sovereign" by nature: regulatory compliance and host security remain the operator’s responsibility.

Measured performance

Compact profile on CPU, full stack, synthetic corpus of 8 documents and 10 questions.

51sFull answer, Compact CPU profile27.8 – 71.3 s
10sClient first token, Compact CPU profile3.9 – 17.9 s
21.5tokens/sGeneration throughput, Compact CPU profile15.8 – 52.0 tokens/s
7.8GBPeak memory, Compact CPU profilereranker up to 5.7 GB; Ollama ≈ 3 GB; steady state (RC2): TEI 3.5 GB, Ollama 3.0 GB

See all measurements and the methodology

Modular architecture

One application, one model gateway, one optional pool. Each component has a role and a contract.

ragatm-open

required

Local-first document Q&A (RAG) platform: isolated document spaces, asynchronous ingestion, hybrid retrieval + reranking, cited answers, models run locally through AI Gateway Open.

View on GitHub

ai-gateway-open

required

Local inference gateway: generation (aggregated and NDJSON streaming), embeddings with verified pipeline identity, cross-encoder reranking, strict model registry, typed errors, non-inferential health probes. No remote provider.

View on GitHub

ai-gateway-pool-open

optional

Standalone HTTP router in front of one or more AI Gateway Open instances, without changing the gateway-api/1.0 contract: node selection, circuit breaker, upstream allowlist, job affinity, saturation handling. Optional for Compact.

View on GitHub

Explore the architecture Open source

Install the Compact profile

Two repositories to clone, a preflight, models to provision, then an offline stack.

Linux x86_64 · Docker Compose · 8 threads · 14 GB RAM · no GPU

git clone https://github.com/ragatm-open/ragatm-open.git
git clone https://github.com/ragatm-open/ai-gateway-open.git   # requis ; placé à côté de ragatm-open/
cd ragatm-open/compact && scripts/preflight.sh

Installation guide

Frequently asked questions

Are my documents sent to an external AI service?

Not in the qualified configuration: all models run locally and the published code contains no fallback to a remote provider.

Do I need a GPU?

No. The Compact profile is qualified on CPU. GPU profiles are described but not qualified.

Is there an official Docker image?

No: images are built locally from source, with base images pinned by digest.

Does it read scanned documents or audio?

Not in the published Open edition: OCR is a disabled build option, and the gateway contains no speech-to-text engine.

All questions