AI Support Bot RAG Demo

Live portfolio demo: https://demo-ai-support-bot.vercel.app

This repo now uses a static Vercel landing demo as the public entrypoint so reviewers do not hit a Streamlit sleep/wake screen. The Streamlit app and Python RAG backend are still included for local technical review.

This is a lightweight proof-of-concept for an AI support assistant that ingests two source types:

• structured and unstructured knowledge base articles
• historical support ticket exports that are cleaned, structured, chunked, embedded, and retrieved

The app answers support questions with grounded citations such as [KB-003] or [TCK-1003], then displays the retrieved sources in a side panel.

1. Demo Architecture

Recommended 48-hour prototype stack:

Layer	Choice	Why
Frontend	Static HTML on Vercel, Flask/Streamlit locally	Reliable public demo link plus richer local technical review
LLM	Gemini API or OpenAI Responses API	Optional API-backed answer synthesis with a provider switch
Embeddings	Local hashed vectors, optional text-embedding-3-small	Lightweight retrieval that works without external services
Retrieval	In-memory cosine search	Keeps the Vercel demo small and reliable
Data cleaning	Python CSV utilities	No heavy data dependency needed for the mock support exports
Deployment	Vercel	Quick client demo link

For a production MVP, this can evolve to managed Pinecone/Qdrant/Chroma, scheduled ingestion, role-based access control, analytics, ticket deflection tracking, and human handoff.

2. Quick Start

python -m venv .venv
.venv\Scripts\activate
pip install -r requirements-local.txt
pip install -r requirements-streamlit.txt
python generate_sample_data.py
python rag_backend.py
streamlit run streamlit_app.py

Copy .env.example to .env and set either GEMINI_API_KEY or OPENAI_API_KEY for LLM-generated answers. Without an API key, the demo still runs with local deterministic embeddings and an extractive fallback answer.

copy .env.example .env

To use OpenAI embeddings instead of local demo embeddings, set:

USE_OPENAI_EMBEDDINGS=true
OPENAI_EMBEDDING_MODEL=text-embedding-3-small
OPENAI_CHAT_MODEL=gpt-5

To use Gemini for answer generation, set:

LLM_PROVIDER=gemini
GEMINI_API_KEY=your_key_here
GEMINI_MODEL=gemini-2.5-flash

Vercel Deployment

This repo includes index.html, an always-awake static landing/demo for Vercel. It shows the support automation workflow, canned grounded answers, retrieved source cards, and an unsupported-answer handoff example without requiring Streamlit or a Python server to wake up.

Live demo: https://demo-ai-support-bot.vercel.app

The Vercel deployment is intentionally static for first-impression reliability. .vercelignore excludes the local Python files from production deploys, while flask_app.py, rag_backend.py, and streamlit_app.py remain in the repo to show the deeper Python RAG implementation and can be run locally.

For the local Flask or Streamlit RAG version, add these environment variables:

LLM_PROVIDER=gemini
GEMINI_API_KEY=your_key_here
GEMINI_MODEL=gemini-2.5-flash

Then run the local version. The public Vercel page does not need API keys.

3. What The Demo Proves

• It ingests both KB articles and historical ticket exports, normalizes each into citation-ready source records, chunks them, embeds them, and retrieves them with in-memory vector search.
• It performs RAG retrieval over both source types and prompts the LLM to answer only from retrieved context with mandatory source IDs.
• It provides a clean chat interface where the answer and the underlying sources are visible together, making hallucination control tangible for the client.

4. Demo Walkthrough Script

• "I built a working mini-version of your AI Support Bot using mock KB articles and historical tickets, because the hard part is not just chat; it is cleaning and structuring messy historical support data so retrieval works."
• "Each response is grounded in retrieved sources and cites the exact KB article or ticket ID, which is how we reduce hallucinations and make the bot usable for both customers and internal support agents."
• "The MVP path is straightforward: replace the mock CSVs with your exports, add scheduled ingestion, run evaluation on historical tickets, and deploy the chat widget behind your existing support workflow."

5. File Map

• generate_sample_data.py: creates mock support ticket and KB CSVs
• rag_backend.py: chunks, embeds, indexes, retrieves, and answers with citations
• index.html: static Vercel landing/demo that avoids Streamlit sleep screens
• streamlit_app.py: local chat UI with source panel and expandable evidence
• flask_app.py: optional Flask chat UI for local/serverless experiments
• requirements-local.txt: Python dependencies for local RAG and Flask review
• .vercelignore: keeps the public Vercel deploy static and always awake
• .env.example: configuration template