Skip to content

Private client/work PoC - sanitized case study · 2026

Customer Support RAG Web Chatbot PoC

AI WorkflowEvaluationPre-Sales

Original customer-support RAG web chatbot PoC with crawl/index/retrieve/generate flow, web demo UI, and honest retrieval-quality evaluation.

Business Problem

A support organization needed a demoable way to answer product-support questions using existing knowledge pages while showing where RAG helps and where retrieval quality still needs work.

Role Scope

Built and framed the web chatbot PoC, including support-content ingestion, retrieval pipeline, answer generation, demo flow, and evaluation notes.

Solution Direction

Used support content ingestion, FAISS/BM25 retrieval, reranking, LLM relevance filtering, answer generation, and a FastAPI web UI to test the support-chat experience.

Deliverables

  • RAG web chatbot
  • Support-content index
  • Evaluation summary
  • Lessons for API follow-up

Impact

Provides the origin story behind the later callbot-compatible API PoC and shows honest handling of measured retrieval limitations.

Description

Private client/work PoC - sanitized case study. Customer-support RAG web chatbot prototype with crawl/index/retrieve/generate flow and evaluation notes.

Source repository is private. Public wording avoids customer names and raw source data.

Output / Proof Signals

FastAPIRAGFAISSBM25RerankingEvaluation

Case Study

Case in a 12-block structure

Overview

Private client/work PoC - sanitized case study. Original customer-support RAG web chatbot prototype with crawl, index, retrieve, generate, web demo, and evaluation artifacts.

Context

This was the web-chat origin story behind the later API-focused callbot PoC. Public wording avoids customer names, raw source data, and private repository links.

Business Problem

A support team needed to test whether existing knowledge pages could support conversational troubleshooting with source links and multi-turn clarification.

Role Scope

Built and framed the web chatbot PoC, including data ingestion, indexing, retrieval, answer generation, demo UI, and evaluation summary.

Solution Direction

Use a RAG chatbot to ingest support content, retrieve candidate documents, generate conversational answers, and expose source links for review.

Architecture

Support content ingestion -> FAISS/BM25 retrieval -> reranking -> LLM relevance check -> answer generation -> FastAPI web UI.

Implementation

The PoC demonstrates a working RAG pipeline and web demo. Latest evaluation notes show retrieval/link quality still needs improvement before production claims.

Demo Scenario

A user asks a product-support question, the chatbot rewrites/searches the query, retrieves likely support documents, and returns an answer with source links.

Deliverables

RAG web chatbot, support-content index, evaluation artifacts, demo flow, and lessons that informed the callbot API follow-up.

Impact

Shows a realistic path from business problem to GenAI support demo, including honest measurement rather than vague 'it works' claims.

Risk & Trade-off

Evaluation artifacts showed useful PoC signal but not production-grade retrieval quality. The right framing is prototype plus improvement roadmap.

Consulting Takeaway

The project is most valuable as proof that I can build, evaluate, and explain an early RAG PoC in stakeholder-friendly terms.