Back to case studies
SaaS4 monthsVoltra

Designing a RAG Pipeline That Actually Works in Production

Voltra wanted to add AI-powered support to their B2B platform, but needed accuracy guarantees that typical chatbot solutions couldn't provide.

voltra-ai-assistant.pixelpathsolutions.in
Voltra AI RAG Pipeline
LLM Orchestration Active
97.2% Confidence
Retrieved Context Docs
API_docs_v3.2_final.pdf (Chunk #812)
Latency
1.4s
[User]:How do I authenticate with a JWT token in the web socket interface?
[Voltra AI]:
To auth via WebSockets, send a JSON payload with the event auth containing your JWT:
{ "event": "auth", "token": "Bearer <JWT>" }
Validated against system specs. Citations: [api_ws.md]
● Generation streaming...

65%

Ticket Deflection

Support tickets resolved without human intervention

1.4s

Response Time

Average time to first token in production

97%

Accuracy

Factual accuracy rate in generated responses

+18pts

CSAT

Improvement in customer satisfaction score

The Business Challenge

The knowledge base contained 200K+ documents across support articles, API docs, and internal runbooks. Off-the-shelf RAG solutions produced hallucinations at an unacceptable rate (23% of responses contained factual errors). The system needed to maintain sub-2-second response times while being honest about uncertainty.

Our Engineering Approach

We built a multi-stage retrieval pipeline with hybrid search (dense + sparse vectors), re-ranking, and a confidence scoring system that routes low-confidence queries to human agents. We also implemented a feedback loop that uses agent corrections to improve retrieval quality over time.

Key Architecture Decisions & Tradeoffs

01

Hybrid search over pure vector similarity

Pure vector search missed keyword-specific queries (error codes, API endpoints). By combining dense embeddings with BM25 sparse retrieval and reciprocal rank fusion, we improved retrieval relevance by 34%.

02

Confidence-based routing over always-generate

Rather than forcing the LLM to always produce an answer, we score retrieval confidence and route uncertain queries to human agents with relevant context pre-loaded. This dropped the hallucination rate from 23% to under 3%.

03

Streaming responses with citation chunking

We stream LLM responses token-by-token but inject citation markers in real-time, linking each claim to its source document. Users can verify any statement with one click.

They didn't just build what we asked for - they challenged our assumptions and shipped something better.

Marcus Rivera

VP of Product, Voltra

Case Details

Client Partner
Voltra
Timeframe
4 months
Scope of Work
Architecture audit, design, full implementation

Case Tech Stack

ReactPythonFastAPIOpenAIPineconeAWSRedis

Ready to build something that matters?

Tell us about your project. We'll respond within 24 hours with an honest assessment of whether we're the right fit.