0xrphl/USCIS-knowledge-base-full-website

A comprehensive dataset of 99,489 content chunks from 4,666 pages on the USCIS website, with pre-computed OpenAI text-embedding-ada-002 embeddings (1536 dimensions).

Built for RAG (Retrieval-Augmented Generation), semantic search, and GraphRAG applications focused on U.S. immigration law and policy.

🔗 GitHub: github.com/0xrphl/USCIS-knowledge-base-full-website
🔥 Scraped with: Firecrawl — The open-source web scraping API for AI
🍎 Visualized with: Embedding Atlas — Interactive embedding visualization

🔬 Embedding Atlas — UMAP Clusters

📊 Dataset Summary

Metric	Value
Total content chunks	99,489
Total embeddings	99,488
Unique URLs scraped	4,666
Embedding model	OpenAI text-embedding-ada-002
Embedding dimensions	1,536
Data quality score	99.99%
Source	uscis.gov
Scraping tool	Firecrawl

📦 Files

File	Rows	Description
data/uscis_content.parquet	99,489	Content chunks with metadata
data/uscis_embeddings.parquet	99,488	OpenAI embedding vectors (1536-dim float32)
data/processed_urls.parquet	4,666	All scraped URLs
data/scraping_metadata.parquet	28	Scraping session logs
data/assessments.parquet	5	Immigration assessments

🔧 Usage

Load with 🤗 Datasets

from datasets import load_dataset

# Load content chunks
content = load_dataset("0xrphl/USCIS-knowledge-base-full-website", "content", split="train")
print(content[0])

# Load embeddings
embeddings = load_dataset("0xrphl/USCIS-knowledge-base-full-website", "embeddings", split="train")
print(len(embeddings[0]["embedding"]))  # 1536

Load with Pandas

import pandas as pd

content = pd.read_parquet("hf://datasets/0xrphl/USCIS-knowledge-base-full-website/data/uscis_content.parquet")
embeddings = pd.read_parquet("hf://datasets/0xrphl/USCIS-knowledge-base-full-website/data/uscis_embeddings.parquet")

Restore into PostgreSQL (Docker)

See the GitHub repo for a full Docker Compose setup that restores this dataset into PostgreSQL + pgvector, Milvus, and Neo4j.

git clone https://github.com/0xrphl/USCIS-knowledge-base-full-website.git
cd USCIS-knowledge-base-full-website
cp .env.example .env
docker compose up -d

📋 Schema

uscis_content — Content Chunks

Column	Type	Description
id	string (UUID)	Unique chunk ID
url	string	Source USCIS URL
title	string	Page title
content	string	Text content (chunk)
created_at	timestamp	When scraped
last_updated	timestamp	Last update
immigration_category	string	General, Humanitarian, Employment, etc.
document_type	string	PDF, Web Page, Policy Manual, Form, News
section	string	USCIS section
source	string	Source identifier
chunk_num	int32	Chunk position (1-indexed)
total_chunks	int32	Total chunks for this URL

uscis_embeddings — Vector Embeddings

Column	Type	Description
id	string (UUID)	Embedding ID
content_id	string (UUID)	FK → uscis_content.id
embedding	float32[1536]	OpenAI text-embedding-ada-002 vector
created_at	timestamp	When generated

📈 Data Breakdown

By Document Type

Type	Chunks	%
PDF	55,957	56.2%
Web Page	32,625	32.8%
Policy Manual	8,886	8.9%
Form	1,156	1.2%
News	800	0.8%
Image	65	0.1%

By Immigration Category

Category	Chunks	%
General	92,520	93.0%
Humanitarian	2,488	2.5%
Employment	2,219	2.2%
Permanent Residence	1,553	1.6%
Citizenship	573	0.6%
Family	136	0.1%

Content Length Statistics

Metric	Characters
Min	3
Average	1,426
Median	1,372
Max	7,999

⚖️ License

Content is sourced from USCIS.gov, a U.S. government website. U.S. government works are in the public domain under 17 U.S.C. § 105.

🙏 Acknowledgments

• Data scraped using Firecrawl
• Embeddings generated with OpenAI text-embedding-ada-002
• Visualized with Embedding Atlas (Apple)