Data Dynamics
Solutions · Open Source

Argus RAG Studio

An open-source, self-hosted platform that covers the full RAG lifecycle — Build, Retrieve & Generate, Evaluate, and Operate & Deploy — in one place. Not a "works-once RAG demo": it ships with an evaluation harness that measures quality in numbers, config sweeps that auto-explore optimal settings, a feedback loop, and agent-based remote deployment, so it runs on-premises and in air-gapped (closed) networks.

Apache License 2.0 · Open SourceGitHub RepositoryProduct Brochure
5 parsing strategies8 chunking strategies5 vector stores

01 · RAG concepts and the whole flow

Start with how RAG uses documents to support answers, then follow the path from preparation to operation.

What is RAG?

RAG retrieves documents relevant to a question and supplies them as context to the language model. Citations accompanying the answer help users check the supporting evidence.

Without RAG vs. with RAG — the indexing path and the query path
Without RAG vs. with RAG — the indexing path and the query pathSelect the diagram to zoom to its original size.

The RAG lifecycle

Prepare searchable knowledge from documents, retrieve evidence for each question and generate an answer. Use evaluations to improve settings, then manage runtime versions, service health and deployment.

Prepare knowledge → answer with evidence → measure and improve → operate and extend
Prepare knowledge → answer with evidence → measure and improve → operate and extendSelect the diagram to zoom to its original size.

02 · Prepare documents for search

Read documents, handle sensitive information, and turn the content into searchable passages and vectors.

Build

Bring in documents, select a knowledge base, then parse, chunk and index

Bring in documents → assign and schedule → prepare searchable content → persist and inspect
Bring in documents → assign and schedule → prepare searchable content → persist and inspectSelect the diagram to zoom to its original size.

Document routing

Routing selects a knowledge base for a document before ingestion. Configure a policy using filename, path, metadata, embedding similarity or LLM judgement, then record the destination and the reason for the decision.

Choose a destination → intake mode → routing rules → decision record and review
Choose a destination → intake mode → routing rules → decision record and reviewSelect the diagram to zoom to its original size.

Document parsing

Parsing extracts usable text and structure from a file. Choose a parser suited to the file type and document layout, then compare the extracted Markdown tables, headings and reading order with the original.

Inspect the input → choose a parser → extract text and structure → review before indexing
Inspect the input → choose a parser → extract text and structure → review before indexingSelect the diagram to zoom to its original size.

A pipeline for Korean documents

Choose processing paths for HWP/HWPX tables, scanned PDFs, multi-column layouts and Korean sentence boundaries. Compare extracted content with the original and review document revisions and image labels before using them in retrieval.

Process and review HWP tables, scanned pages, Korean sentences, document revisions and image labels
Process and review HWP tables, scanned pages, Korean sentences, document revisions and image labelsSelect the diagram to zoom to its original size.

PII masking

PII is personally identifiable information. When enabled, transforms mask text matching configured rules after parsing and before chunking and embedding. Check for missed detections and manage original-file storage and access separately.

Transform placement → before and after masking → detection rules → errors and original files
Transform placement → before and after masking → detection rules → errors and original filesSelect the diagram to zoom to its original size.

Chunking

Chunking divides a document into retrievable passages. Large chunks can mix unrelated content; small chunks can lose needed context. Choose boundaries, size and overlap for the document, then check the retrieval results.

Split into chunks → choose boundaries → adjust size and overlap → preview and evaluate
Split into chunks → choose boundaries → adjust size and overlap → preview and evaluateSelect the diagram to zoom to its original size.

Embedding — into vectors retrieval can match

The embedding settings define a collection's vector space. Provider, model, dimension and distance metric are fixed when the collection is created, and changing them later means reindexing. On closed networks or in CI, the wiring can be verified with no embedding server at all.

How chunks and questions become vectors and are compared in a compatible model space
How chunks and questions become vectors and are compared in a compatible model spaceSelect the diagram to zoom to its original size.

03 · Find evidence and generate answers

Find relevant evidence in the prepared documents, rank it, and use it to generate answers with citations.

Retrieve & Generate

Find relevant evidence and generate an answer with citations

From parallel vector and lexical search to a cited answer
From parallel vector and lexical search to a cited answerSelect the diagram to zoom to its original size.

Retrieval — how the evidence is found

The question feeds vector and lexical search, and RRF combines their rankings. Collection scope and metadata filters narrow the results; PostgreSQL supplies the canonical chunk text and metadata.

Search scope → vector and lexical candidates → canonical text and RRF → evidence chunks
Search scope → vector and lexical candidates → canonical text and RRF → evidence chunksSelect the diagram to zoom to its original size.

Reranking — search wide, select narrow

Reranking reassesses retrieved candidates against the question and changes their order. It cannot recover missing candidates, and changing its settings does not require reindexing.

Candidate chunks → question–passage assessment → reorder and select → reranker settings
Candidate chunks → question–passage assessment → reorder and select → reranker settingsSelect the diagram to zoom to its original size.

From internal documents to cited responses

Separate document indexing from question handling, then provide retrieved evidence as generation context.

Four-stage architecture for document indexing, hybrid retrieval, RRF/reranking and cited generation
Prepare documents → build index → retrieve evidence → generate citationsSelect the diagram to zoom to its original size.
1.Prepare documents
Route to a collection, parse the document and apply enabled PII transforms.
2.Build the index
Store chunks, metadata and vectors while keeping original documents in object storage.
3.Retrieve evidence
Fuse vector and lexical search ranks with RRF, then rerank when configured.
4.Generate with citations
Pass retrieved passages to the model and map citation markers in streamed answers to chunks.

PII transforms and external inference servers are optional. Citation markers map to retrieved chunks; this does not guarantee the accuracy of every answer sentence.

View GitHub documentation and code

04 · Evaluate and improve quality

Check retrieval and answer quality, compare configurations, and promote a validated version to production.

Evaluation — quality as a number

Retrieval is scored without an LLM through Hit Rate and MRR; answers are scored by an LLM judge on faithfulness, relevance and correctness. Configuration sweeps split the search space by whether reindexing is needed, and a holdout split with an overfit flag validates the winner.

Reference questions and sources, a worked Hit Rate/MRR example, LLM assessment and holdout comparison
Reference questions and sources, a worked Hit Rate/MRR example, LLM assessment and holdout comparisonSelect the diagram to zoom to its original size.

Evaluate & Operate

Evaluate quality and manage settings, runtime versions and operational status

The measure → optimise → improve loop
The measure → optimise → improve loopSelect the diagram to zoom to its original size.

Turn evaluation results into operating settings

Measure quality on golden datasets, compare configurations, then manage pipeline versions and service deployment.

Four-stage loop for baseline evaluation, configuration sweeps, pipeline versioning and agent deployment
Evaluate → compare settings → select and version → deploy and observeSelect the diagram to zoom to its original size.
1.Evaluate a baseline
Record Hit Rate, MRR and LLM judge metrics using golden questions.
2.Compare configurations
Review index/query sweep results, holdout scores and overfitting flags.
3.Select and version
Review and apply chosen settings, then save runtime settings as a pipeline version.
4.Deploy and observe
Deploy workers and inference services through host agents; feed traces and feedback into the next evaluation.

Sweeps compare configurations within the selected search space. Index changes may require reindexing; applying recommended query settings and deploying services are separate actions.

View GitHub documentation and code

05 · Understand the system

See how services and storage divide the document processing, retrieval, and evaluation work introduced above.

Concept diagram

The API and workers connect document preparation with question processing. PostgreSQL stores chunks and operational data, S3/MinIO stores originals and images, and inference services and host agents support processing and deployment.

Argus RAG Studio system overview
Argus RAG Studio system overviewSelect the diagram to zoom to its original size.

06 · Deploy and extend operations

With the core flow in place, explore air-gapped environments, agents, integrations, and operational features.

Deployment in an air-gapped network

Import container images and model packs separately, then deploy workers and inference services through each host’s Argus Agent. Offline operation requires internal registries, model storage and local endpoints.

Remote agent deployment and the air-gapped model supply path
Remote agent deployment and the air-gapped model supply pathSelect the diagram to zoom to its original size.

Improve quality and operate in closed networks

Connect evaluation, retrieval training, host agents and image review to the existing RAG pipeline.

Improve quality and operate in closed networks — Evaluation and versions · Retrieval training · Closed-network deployment · Image workflows
Connect evaluation, retrieval training, host agents and image review to the existing RAG pipeline.Select the diagram to zoom to its original size.
1.Evaluation and versions
Compare holdout metrics, apply index changes and version runtime settings.
2.Retrieval training
Review synthetic triplets and use external training and model registration.
3.Closed-network deployment
Import images and models separately and deploy services through host agents.
4.Image workflows
Review OCR annotations, inspect images and index VLM-derived content.

Offline use needs prepared artifacts and local endpoints. Models are separate from container images; changing embeddings may require reindexing.

View GitHub documentation and code
Apache License 2.0 · Open Source

An open-source RAG platform

Argus RAG Studio is published on GitHub under the Apache License 2.0. The entire RAG engine — backend (FastAPI), frontend (Next.js), and the standalone embedding/reranker servers — is open, so enterprises can verify the code directly, extend it to fit their environment, and operate it without sending data outside.

  • Apache 2.0 with no commercial-use restrictions
  • Verify and extend the code yourself
  • Self-host in air-gapped / on-premises
GitHub RepositoryContact us
Contact usProduct Brochure