RAG/Case StudyShipped/ 2026

Atlas RAG

Knowledge-aware retrieval over private docs.

PythonFastAPIVector DBLangChain

/atlas-rag

atlas rag / onlinePython · FastAPI · Vector DB · LangChain

01/Problem

Most RAG systems either lose their grounding on long documents or surface confident-but-wrong answers with no way to tell the two apart. Knowledge teams needed something they could actually put in front of customers.

02/Solution

Atlas RAG combines hybrid retrieval (BM25 + dense), a cross-encoder reranker, and a citation-locked generator that refuses to answer when retrieval confidence is low. Every response ships with line-level citations and a calibrated confidence score.

03/How it works

The system, end to end.

01next →
Query
User question + optional filters.
02next →
Retrieve
Hybrid BM25 + dense recall, top-50.
03next →
Rerank
Cross-encoder narrows to top-8.
04next →
Compose
Citation-locked prompt assembly.
05loop
Generate
Grounded answer + citations + score.

04/Architecture

Components and data flow.

How the moving parts fit together. Solid arrows are synchronous calls; dashed arrows are queued or batched.

01→
Client (web/SDK)
02→
API gateway
03→
Query router
04→
Hybrid retriever (BM25 + dense)
05→
Reranker (cross-encoder)
06→
Citation-locked LLM
07→
Postgres + vector store
08
Telemetry pipeline

↓ flow continues

05/Technology

The stack.

AI / ML: OpenAI / Anthropic modelsBGE embeddingsCross-encoder reranker
Frontend: Next.jsTailwind
Backend: PythonFastAPILangChain
Database: PostgreSQLpgvectorRedis
Infrastructure: GitHub Actions
APIs: OpenAIAnthropicCohere rerank

06/Engineering decisions

What we picked, and why.

01
Hybrid retrieval instead of dense-only
Choice
BM25 + dense with reciprocal-rank fusion
Alternatives
- Dense-only retrieval
- Graph retrieval
Why
Lexical recall catches identifiers and rare terms that dense embeddings flatten; fusion keeps both signals honest.
02
Citation-locked generation
Choice
Force the model to cite, refuse when unsupported
Alternatives
- Free-form generation + post-hoc citations
Why
Locking citations into the decoding loop removed the 'confident but ungrounded' failure mode we saw in early evals.
03
pgvector instead of a dedicated vector DB
Choice
Postgres + pgvector
Alternatives
- Pinecone
- Weaviate
- Qdrant
Why
Customers already ran Postgres. One fewer system to operate, transactional consistency for free.

07/Evaluation

What we measured.

01
94%
Grounded
02
91%
Citation precision
03
1.4s
p95 latency
04
6%
Refusal rate

08/Challenges

What almost broke us.

01
Long-document retrieval drift
What happened
Answers slowly drifted as corpus grew.
What we did
Introduced per-chunk metadata anchors + cross-encoder reranker.
02
Citation hallucination
What happened
Model cited plausible-but-wrong passages.
What we did
Switched to citation-locked decoding and required refusal when no support.

09/Results

Reduced ungrounded answers from ~22% to ~6% on the internal eval set. Three enterprise teams moved from pilot to production within a quarter.

10/Future improvements

›Multimodal retrieval over diagrams and tables
›Streaming answers with progressive citations
›Tenant-aware evals as a managed service

/explore the build

GitHub Live Demo

Have an interesting AI problem?

Let's build something.

If you're working on something that needs real AI engineering — not just a demo — let's talk.

Start a Conversation

Atlas RAG

Problem and solution

Hybrid retrieval instead of dense-only

Citation-locked generation

pgvector instead of a dedicated vector DB

Have an interesting AI problem?