StayChat AI — RAG-Based Hotel Q&A System

Retrieval-Augmented Generation pipeline for answering natural-language questions about hotels from a curated document corpus (40 synthetic documents).

Submission guide: see SUBMISSION.md — add your name before submitting.

Quick start (clone from GitHub)

git clone https://github.com/YOUR_USERNAME/YOUR_REPO.git
cd YOUR_REPO
python -m venv .venv
.venv\Scripts\activate          # Windows
pip install -r requirements.txt
python main.py --all

The FAISS index is not in the repo (see .gitignore); main.py --all rebuilds it automatically.

Architecture

flowchart LR
  A[hotel_documents.json] --> B[Preprocess and Chunk]
  B --> C[MiniLM Embeddings]
  C --> D[FAISS Index]
  B --> E[BM25 Index]
  F[User Query] --> G[Hybrid Retrieve]
  D --> G
  E --> G
  G --> H[Rerank]
  H --> I[Top-k Chunks]
  I --> J[LLM OpenAI Ollama or Mock]
  I --> K[Hallucination Controls]
  K --> J
  J --> L[Answer with Citations]

Loading

Component	Choice	Rationale
Chunking	Sentence-aware, 280 chars, 60 overlap	Splits longer amenity/policy passages; overlap preserves WiFi/breakfast context across boundaries
Embeddings	all-MiniLM-L6-v2 (384-d)	Strong semantic search, local, no API cost
Vector DB	FAISS IndexFlatIP + BM25 hybrid	Dense recall + sparse keyword match (e.g. “complimentary breakfast”)
Top-k	7 (10 for list queries)	Better multi-hotel recall than k=5
LLM	OpenAI / Ollama / generic mock	Mock is query-agnostic (no hardcoded demo branches)
Hallucination	Threshold + term grounding + prompt + verification	See outputs/hallucination_ablation.md

Dataset (40 documents)

Category	Count
Hotel descriptions	9
Amenities	7
Guest reviews	10
Policies	7
Location	7

Source: synthetic (MIT). File: data/hotel_documents.json.

Project layout

Task/
├── SUBMISSION.md
├── config.py
├── main.py
├── requirements.txt
├── data/
├── src/
│   ├── preprocessing.py
│   ├── embed_store.py
│   ├── hybrid_retrieval.py
│   ├── retrieval.py
│   ├── generation.py
│   ├── evaluation.py
│   └── hallucination.py
├── tests/test_rag.py
├── index/
└── outputs/

Setup

cd Task
python -m venv .venv
.venv\Scripts\activate
pip install -r requirements.txt

Optional:

set OPENAI_API_KEY=sk-...
set USE_OLLAMA=1

Run

python main.py --all
python -m unittest tests.test_rag -v

Example queries

ID	Query
Q1	Which hotels have free WiFi and complimentary breakfast?
Q2	What is the cancellation policy of Hotel X?
Q3	Suggest a hotel with excellent reviews near the beach.

Results: outputs/sample_outputs.md

Known limitations

1. Mock LLM uses lexical overlap — enable OpenAI or Ollama for fluent paraphrasing.
2. Manual relevance labels for three queries; metrics are indicative.
3. Production would add cross-encoder reranking and NLI-based verification.

Task checklist

• Task 1: Cleaning, chunking (280/60) with justification
• Task 2: Hybrid embeddings + FAISS/BM25 + top-k
• Task 3: Context-only prompt + Q1–Q3
• Task 4: Metrics with workings + detailed qualitative analysis
• Task 5: Hallucination control + ablation