Building software for clients since 2010
- 15+
- Years shipping software
- 84%
- Retainer renewal rate
- 4–8
- Weeks to first production ship
AI · AI & Technology
The making of RAG Knowledge Assistant
Customer-facing RAG chatbot with hybrid search, reranking, and cited answers from 50K+ product documents — handling 80% of tier-1 support tickets autonomously.
- Client
- Enterprise Client
- Timeline
- 8 weeks
- Built with
- Claude API, Vercel AI SDK, pgvector, Next.js, Cohere
Challenges
- Ensuring answer accuracy across 50K documents
- Low-latency retrieval at scale
- Handling ambiguous queries gracefully
Solution
We developed a customer-facing RAG chatbot that uses hybrid search and reranking to provide accurate, cited answers from a knowledge base of 50,000+ product documents. The system handles tier-1 support tickets autonomously with human escalation for complex queries.
Services provided
Colors
#5C61A4
#2153BC
#22254F
#3E417D
#9599D9
What we built
Hybrid Search with pgvector
Combines dense vector similarity and sparse keyword matching for retrieval that handles both semantic and exact queries.
Cohere Reranking for Relevance
Two-stage retrieval with neural reranking to surface the most relevant documents from thousands of candidates.
Cited Answers with Source Links
Every response includes clickable citations linking back to the exact source documents for full transparency.
Human Escalation Workflow
Intelligent routing that detects complex or sensitive queries and seamlessly hands them off to human agents.
Analytics Dashboard
Comprehensive metrics on query volumes, resolution rates, user satisfaction, and knowledge gap identification.
Multi-Language Support
Natural language understanding and response generation in multiple languages for global customer bases.
Results
80%
of tier-1 tickets handled autonomously
3.2s
average response time
92%
user satisfaction score
