README.md

---
license: mit
task_categories:
  - text-retrieval
  - question-answering
language:
  - en
tags:
  - building-defects
  - construction-quality
  - ncc
  - vba
  - rag
  - embeddings
  - bge
pretty_name: Building Defects RAG Index
size_categories:
  - 1K<n<10K
---

# Building Defects RAG Index

This dataset contains pre-built retrieval artifacts for the Building Defects
& Quality chatbot. It includes chunked documents, embeddings, and search
indexes ready for deployment.

## Purpose

This dataset enables Retrieval-Augmented Generation (RAG) for answering
questions about building defects, construction tolerances, NCC Volume Two
requirements, and VBA Guide to Standards and Tolerances.

## Embedding Model

- **Model**: `BAAI/bge-small-en-v1.5`
- **Dimension**: 384
- **Storage**: float16 for 50% compression

## Dataset Schema

### chunks.parquet

Document chunks with metadata for retrieval.

| Column | Type | Description |
|--------|------|-------------|
| chunk_id | string | Unique identifier for the chunk |
| text | string | The chunk text content |
| heading_path | list[string] | Hierarchical heading context |
| source | string | Source PDF filename |
| page | int | Page number in source document |
| chunk_hash | string | SHA-256 hash (first 16 chars) |

### embeddings.parquet

Embedding vectors for semantic search.

| Column | Type | Description |
|--------|------|-------------|
| chunk_id | string | Matches chunk_id in chunks.parquet |
| chunk_hash | string | SHA-256 hash for deduplication |
| embedding | fixed_size_list[float16] | Embedding vector |

### faiss_index.bin

Serialized FAISS index for fast approximate nearest neighbor search.
Uses IndexFlatIP (inner product) for cosine similarity search.

### bm25_index.pkl

Serialized BM25 index for lexical/keyword search.
Used in hybrid retrieval with reciprocal rank fusion.

### source_manifest.json

Manifest of source files with hashes for change detection.

| Field | Type | Description |
|-------|------|-------------|
| sources | list[SourceFile] | Source file metadata |
| created_at | string | ISO 8601 creation timestamp |
| total_chunks | int | Total chunk count |
| total_embeddings | int | Total embedding count |

### index_version.txt

Single-line version identifier for cache invalidation.
Format: `{timestamp}_{hash}` (e.g., `20240115_123456_abc123`)

## Statistics

- **Total Chunks**: 1,523
- **Total Embeddings**: 1,523
- **Embedding Dimension**: 384
- **Source Documents**: 0
- **Storage Format**: Parquet with Snappy compression

## Usage

### Loading with HuggingFace Datasets

```python
from datasets import load_dataset

# Load chunks
chunks = load_dataset("sadickam/BuildingDefect_index", data_files="chunks.parquet")

# Load embeddings
embeddings = load_dataset("sadickam/BuildingDefect_index", data_files="embeddings.parquet")
```

### Loading with huggingface_hub

```python
from huggingface_hub import hf_hub_download

# Download FAISS index
faiss_path = hf_hub_download(
    repo_id="sadickam/BuildingDefect_index",
    filename="faiss_index.bin",
    repo_type="dataset",
)

# Download BM25 index
bm25_path = hf_hub_download(
    repo_id="sadickam/BuildingDefect_index",
    filename="bm25_index.pkl",
    repo_type="dataset",
)
```

### Loading FAISS Index

```python
import faiss

index = faiss.read_index(faiss_path)
```

## License

This dataset is released under the MIT License.

## Sources

This dataset is based on:
- **NCC Volume Two (2022)** - National Construction Code, Australian Building Codes Board
- **VBA Guide to Standards and Tolerances** - Victorian Building Authority