Frequently asked questions
Short answers, checked against the published code. Detailed pages are linked where they exist.
Understanding
What does "RAG against the machine" mean?
RAG stands for Retrieval-Augmented Generation: the model answers from passages retrieved in your documents, not from its memory alone. "Against the machine" recalls the project’s stance: keep your documents and models on your own infrastructure. The name has no connection with the band.
What does RAGatm Open actually do?
You upload documents into an isolated space, the application chunks and indexes them, then you ask questions in plain language. The answer is generated by a local model and cites the passages used (document, page or slide, excerpt).
Are the answers reliable?
They are source-backed: every answer points to the retrieved passages, which you can open. But a language model can be wrong or summarise poorly; check the cited source before any decision. The Compact profile model (1.5 billion parameters) demonstrates the mechanism; its documentary quality is not claimed.
Does RAGatm Open give legal advice?
No. It retrieves and summarises passages from your documents. It provides no legal advice, regulatory analysis or compliance guarantee.
Data and sovereignty
Are my documents sent to an external AI service?
Not in the qualified configuration: all models run locally behind AI Gateway Open, the only accepted value for the external-provider fallback is none, and startup checks reject any other value. The Compact stack’s internal networks have no Internet egress in nominal operation.
Is it GDPR-compliant?
The software helps you control data (local hosting, local models, controlled backups), but the compliance of a processing activity depends on the operator: purposes, retention, access rights, host security. No software is "compliant by default".
Who can access what?
Users belong to workspaces with roles (owner, admin, member, viewer); documents live in document spaces isolated from one another. Every account must enrol a TOTP second factor; there is no public sign-up.
Installing and running
What hardware is needed?
The Compact profile is qualified on an x86_64 machine with 8 threads and 14 GB of RAM, no GPU. Expect about 3 GB of application images and 4 GB of models, plus your data. A GPU is not required; GPU profiles are not qualified. See the Installation page.
Is there a ready-made Docker image?
No. No official image is published: images are built locally from source, with base images pinned by digest. There is no one-command install.
Do I need an Internet connection?
Only while building images and provisioning models. Afterwards the stack runs offline; a script verifies the absence of network egress.
Which models are used?
In the Compact profile: granite-embedding:278m (embeddings), qwen2.5:1.5b-instruct (generation) and BAAI/bge-reranker-v2-m3 (reranking), all Apache-2.0 licensed, downloaded by the operator. Changing the embedding model requires reindexing; no implicit reindexing happens.
How fast does it answer?
On the CPU qualification workstation, a full answer takes between 28 and 71 s (median ≈ 51 s), including about 10 s of reranking. GPU profiles are not qualified, and reranking runs on CPU in every current profile. Details and caveats on the Performance page.
How many simultaneous users?
The gateway accepts two concurrent generations, then answers 503 with a retry delay. Load beyond one concurrent generation is not qualified: the Compact profile targets a workstation or a small team.
Formats and capabilities
Which formats are supported?
By default in Compact: text, Markdown, PDF with a text layer, DOCX, CSV, HTML, XLSX and PPTX. XML, DXF and EML ship but are closed by a flag. Images, scanned PDFs (OCR), IFC, DWG and audio require an image rebuild or a missing component. The full catalogue details 44 formats.
Does it read scanned documents?
Not in the published images: OCR (local Tesseract) is a build option disabled in Compact. A scanned PDF is accepted at upload but its ingestion fails with an explicit code.
Does it understand drawings and diagrams?
No. The DXF extractor (and DWG through conversion) indexes text annotations, attributes and layers; it reconstructs neither the geometry nor the meaning of a drawing. No vision model is enabled in the Open edition.
Can it transcribe audio recordings?
The audio path exists in the application code, but AI Gateway Open contains no speech-to-text engine. Without an operator-provided ASR service, audio remains unavailable.
Project
Under which licence?
Apache License 2.0 for the three repositories; third-party components keep their licences (listed in each repository). See Open source.
What is the Pool?
AI Gateway Pool Open is an optional HTTP router placed in front of one or more gateways. It relays the same API contract, manages per-node capacity and transport failures, and performs no inference. It is absent from the Compact profile.
How do I report a vulnerability?
Privately, as described in SECURITY.md (contact@ragatm.com) or through GitHub private vulnerability reporting.