Multi-Intent Automobile Chatbot with RAG

Project Overview

The project focused on building a multi-intent AI chatbot for an automobile client with a large knowledge base (~100 GB of car manuals). The chatbot was designed not only to answer user questions through Retrieval-Augmented Generation (RAG) but also to handle service scheduling and service offers discovery.

The solution combined custom query routing, RAG with Gemini 2.0 Flash, web scraping from authorized dealers, and Firestore-based session management to deliver a complete conversational assistant for car owners and service seekers.

---

Problem Statement

Automobile customers often require quick access to diverse types of information:

• Car details: Looking up information from lengthy car manuals.
• Service scheduling: Booking or inquiring about car servicing.
• Promotional offers: Finding latest offers from authorized dealers.

Traditional chatbots fail in this domain because they:

• Can’t handle multiple intents beyond FAQ-style responses.
• Struggle with large technical documents (hundreds of GBs).
• Don’t integrate seamlessly with lead generation workflows or dealer offers.
• Don’t restrict results to only the user’s car manuals, which causes irrelevant answers.

The challenge was to create a scalable, multi-functional chatbot capable of intent recognition, retrieval-based answering, and lead generation in a unified system.

Objectives

1. Implement a query routing mechanism to identify user intent (manual lookup, service scheduling, offers).
2. Deploy a RAG pipeline to provide accurate, context-aware answers from 100 GB+ of car manuals.
3. Enable service scheduling workflows by collecting user inputs (mileage, issues, preferred service).
4. Integrate authorized dealer offer scraping for real-time promotional updates.
5. Store user sessions and histories for contextual continuity.

Methodology

1. Query Routing & Intent Classification

• Built a custom query router to classify queries into three primary routes:
  1. Car Information (RAG)
  2. Service Scheduling
  3. Service Offers
• Router uses LLM-based classification to ensure correct intent mapping before passing the query downstream.

2. Retrieval-Augmented Generation (RAG)

• Knowledge Source: ~100 GB of car manuals across multiple car models.
• Access Control: Each user is restricted to only their car’s manuals.
• Filtering: Before retrieval, the system filters the vector database by car model to ensure only relevant manuals are queried.
• Parsing: Llama Parser used for document reading.
• Vector Database: Vertex AI Vectorstore for embedding storage and retrieval.
• LLM: Gemini 2.0 Flash for response generation.

Workflow:

1. User query received.
2. Router classifies query as manual lookup.
3. Embedding retrieval from Vertex AI Vectorstore.
4. Gemini 2.0 Flash generates contextual, accurate answer.

3. Service Scheduling Flow

• If intent = service scheduling:
  • Chatbot collects user details interactively:
    • Car mileage
    • Issues with the car/ Service type requested
    • Additional issues (If any)
    • if user needs transportation
  • Data stored in Firestore as a lead.
  • Confirmation requested from user before final submission.
  • Simultaneously, chatbot suggests relevant authorized dealer service offers if available.

4. Service Offers Integration

• Implemented scraping pipeline for authorized dealer websites to fetch latest service offers.
• Offers integrated into chatbot responses dynamically:
  • During service scheduling (cross-sell).
  • On direct user request (“Show me offers for my car”).

5. Session Management

• Used Firestore to store:
  • User sessions.
  • Session history (previous queries, scheduling data).

Results

• Accuracy: RAG pipeline delivered high-quality, context-specific answers from only the user’s car manuals.
• Efficiency: Query routing reduced irrelevant responses by ensuring correct flow assignment.
• Lead Generation: Automated service scheduling captured structured leads directly into Firestore.
• User Experience: Integrated offers from authorized dealers improved engagement and upselling opportunities.

Applications

• Customer Support: Provide instant answers from car manuals.
• Lead Generation: Collect structured data for service centers.
• Knowledge Access: Simplify navigation through complex car manuals.

Key Learnings

1. RAG alone isn’t enough — successful domain chatbots require multi-intent routing.
2. Combining structured workflows (service scheduling) with unstructured retrieval (manual lookup) provides a holistic user experience.
3. Session persistence is critical for multi-turn conversations and lead generation.

Conclusion

This project delivered a comprehensive automobile chatbot that went beyond traditional RAG. By combining custom intent routing, Gemini 2.0 Flash RAG with car manual filtering, service scheduling workflows, authorized dealer offer scraping, and Firestore session management, the system enabled the client to:

• Automate service lead generation.
• Improve customer support accuracy.
• Enhance engagement through real-time dealer offers.

The result is a scalable, domain-specific intelligent assistant that addresses multiple customer needs in a single conversational interface.