UCSD Library Chatbot - Agentic RAG & API System

A GenAI chatbot designed to support UC San Diego librarians by asynchronously answering student questions through an agentic two-path workflow: RAG retrieval for knowledge-based queries and API integration for catalog searches.

System Overview

Architecture: Next.js web application with intelligent query routing

Two Processing Paths:

1. RAG Pipeline - LibAnswers FAQ retrieval via Qdrant vector DB + Cohere reranking
2. API Integration - Query reformulation + Ex Libris Primo VE catalog search + reasoning

Core Technologies: OpenAI (routing & generation), Qdrant (vector DB), Cohere (reranking)

Core Components

• Query Router - Routes to RAG or API path
• RAG Pipeline - Semantic search + reranking for FAQs
• API Module - Reformulates queries for Primo VE
• Reasoning Model - Synthesizes catalog results
• Response Generator - Assembles final student-facing answers

Data Sources

LibAnswers FAQ Database

• Library policies & procedures
• Computing resources how-to guides
• Billing information
• Building details

Primo VE Discovery Platform

• Local and consortia library resources

Data Processing

FAQ Preparation

• Consolidate question + details fields
• Strip HTML from answer
• Remove unnecessary fields (owner, short_answer)
• Preserve: questions, answers, topics, keywords, links

Primo VE Processing

1. Reformulate natural language → structured discovery query
2. Restructure API response (title, author, description, availability)
3. Pass clean data to reasoning model

Chunking Strategy

Current Approach: No chunking (FAQ entries sufficiently concise)

Fallback Plan: Semantic chunking for entries >700 words (pending validation)

Data Schemas

FAQ Schema (LibAnswers)

interface FAQ {
  faqid: number;
  group_id: number;
  question: string;
  details: string;
  answer: string;
  topics: Array;
  keywords: Array;
  url: {
    public: string;
    admin: string;
  };
  totalhits: number;
  created: string;
  updated: string;
  votes: {
    yes: number;
    no: number;
  };
  links: Array;
  files: Array;
}

Primo VE API Response Schema

interface PrimoVEResponse {
  resDetails: {
    totalResultsLocal: number;
    totalResultsPC: number;
    total: number;
    first: number;
    last: number;
  };
  docInfo: Array;
}

Processed Primo VE Document (for Reasoning Model)

interface ProcessedPrimoDoc {
  title: string;
  type: string;
  publisher?: string;
  subjects: string[];
  contents?: string;
  permalink: string;
  recordId: string;
  hasFullText: boolean;
}

---

Goal: Reduce librarian workload while delivering accurate, context-aware responses to diverse student inquiries.

Prerequisites

Before getting started, you'll need to set up the following services:

Required API Keys

1. OpenAI API Key (https://platform.openai.com/api-keys)

   • You'll need at least $5 in credits on your OpenAI account
   • Used for embeddings, chat completions, routing logic, and reasoning

2. Qdrant API Key (https://cloud.qdrant.io/)

   • Free tier available
   • Used for vector database storage and semantic search

3. Cohere API Key (https://cohere.com/)

   • Free tier available
   • Used for reranking retrieved FAQ results

4. Ex Libris Primo VE API Key (https://developers.exlibrisgroup.com/)

   • Institutional access required
   • Used for catalog search and resource discovery

5. Springshare LibAnswers API Key (https://springshare.com/libanswers/)

   • Institutional access required
   • Used for retrieving FAQ data from LibAnswers

6. Helicone API Key (https://www.helicone.ai/)

   • Free tier available
   • Used for LLM observability and monitoring

Create a .env file in the root directory with these keys:

OPENAI_API_KEY=your_openai_key_here
QDRANT_API_KEY=your_qdrant_key_here
QDRANT_URL=your_qdrant_cluster_url
COHERE_API_KEY=your_cohere_key_here
PRIMO_VE_API_KEY=your_primo_api_key_here
LIBANSWERS_API_KEY=your_libanswers_key_here
HELICONE_API_KEY=your_helicone_key_here
OPENAI_FINETUNED_MODEL=your_finetuned_model_id (optional)