RAG Chatbot Multi-Tenant AI Chatbot Platform
From content to deployed AI chatbot in minutes a production-grade RAG platform that lets businesses train and embed chatbots on their own data, without writing AI code.
About the client
RAG Chatbot Platform
Remote / SaaS
A production-ready, multi-tenant Retrieval-Augmented Generation platform built for SaaS companies, agencies, and enterprise teams. It enables any business to deploy AI chatbots grounded in their own content with no AI engineering required. Organisations ingest their documentation, configure bot behaviour, and embed a fully branded chat widget on any website, backed by a pipeline that actively prevents hallucination and monitors answer quality.
Technologies used
Results at a glance
5-stage
RAG pipeline eliminating hallucination by design
13+
Quality metrics scored and stored per conversation turn
2
Script tags to embed a fully branded chatbot on any site
Full
Retrieval traceability from query rewrite to final answer
RelayRAG launched as a production-ready platform: businesses go from raw content to a deployed, branded AI chatbot in minutes, with a retrieval pipeline that actively prevents hallucination, automated quality monitoring that flags issues before they reach users, and complete observability across every conversation turn.
Overview
The context
Most businesses that want an AI assistant face the same wall: generic LLMs hallucinate, fine-tuning is expensive, and building a retrieval pipeline from scratch requires significant engineering effort. Agencies and SaaS operators face an additional layer they need to support multiple clients from a single platform while keeping data, configuration, and branding completely isolated.
RelayRAG was built to eliminate all of those blockers at once: a unified platform where teams go from raw content to an embedded, production-grade AI chatbot without touching AI infrastructure and where every answer can be traced, audited, and quality-scored after the fact.
The challenge
The Challenge
Deploying a reliable AI chatbot is harder than it looks. Generic language models fabricate answers with confidence, and without a retrieval layer grounded in real company content, every response is a liability. Fine-tuning solves accuracy but not freshness source content changes faster than a model can be retrained, and the cost of each retraining cycle makes it impractical for most teams.
For agencies and SaaS operators, the multi-tenancy problem compounds everything. Each client needs isolated data, independent bot configuration, and custom branding while the platform operator needs a single operational view across all of them. Building that from scratch, alongside a production-grade retrieval pipeline and post-deployment observability, is an engineering project measured in months.
Our solution
What We Built
A multi-tenant SaaS platform where any team can ingest URLs or documents, configure a chatbot's personality and appearance, and embed it on any website with two script tags. The retrieval pipeline goes well beyond basic vector search: every user query is rewritten and expanded into multiple sub-queries, each is searched independently against pgvector, candidates are deduplicated and LLM-reranked, and the final answer is generated exclusively from verified context with citations and related follow-up questions included automatically.
Every interaction is analysed asynchronously after generation scoring groundedness, faithfulness, hallucination risk, citation accuracy, sentiment, intent, and PII presence with all results visible in an admin dashboard that shows session-level and turn-level detail. Web ingestion runs as background jobs with pause, resume, and reingest support, while full prompt configurability across every pipeline stage means operators are never locked into default behaviour.
- Multi-stage RAG pipeline: query rewriting, sub-query expansion, parallel vector retrieval, LLM reranking, and grounded answer generation
- Automated quality analysis: groundedness, hallucination risk, citation accuracy, sentiment, intent, and PII detection all stored per turn
- Embeddable chat widget with full branding and theme controls no rebuild required
- Web ingestion with pause, resume, reingest, real-time progress tracking, and per-URL logging
- Full conversation observability: session metadata, turn-level retrieval trace, latency, and token usage
- Per-bot prompt configurability across every pipeline stage
- Multi-tenant architecture with tenant-scoped data and role-based access control
- Admin dashboard with platform metrics, analytics, and advanced filtering by tenant, bot, and date
- Streaming-first chat via Server-Sent Events with Time-To-First-Token measurement
- JWT authentication, password reset, and user management
Outcomes
The results
RelayRAG launched as a production-ready platform: businesses go from raw content to a deployed, branded AI chatbot in minutes, with a retrieval pipeline that actively prevents hallucination, automated quality monitoring that flags issues before they reach users, and complete observability across every conversation turn.
5-stage
RAG pipeline eliminating hallucination by design
13+
Quality metrics scored and stored per conversation turn
2
Script tags to embed a fully branded chatbot on any site
Full
Retrieval traceability from query rewrite to final answer
Ready?
Want results like these?
We take on a limited number of engagements. Tell us what you're building.