We are sunsetting our office from Canada and shifting to Hong Kong. Stay connected for latest updates; physical meetings will resume soon.
Platform Pipelines

Engineered Ingestion Architecture.

Explore Reworkd's high-speed crawling clusters, locality-sensitive hashing deduplication filters, and semantic vector formatting pipelines built specifically for model pretraining.

01 / CRAWLER WORKERS

High-Throughput Scrapers

Reworkd operates a decentralized cluster of headless crawling nodes dynamically shifting user agents, viewport scopes, and request throttling configurations to emulate human interactions.

When encountering complex client-side applications (React, Angular, Next.js), we deploy a headless Chromium context (Playwright engine) that waits for async API calls to resolve and compile selectors into structured DOM representations, auto-throttling requests inline to prevent IP rate-limiting.

  • Rotating Proxy subnets
  • Automated Header spoofing
  • CNN Visual slider solver
  • Dynamic layout compilations
LIVE WORKER SIMULATOR
[04:39:00] [SYSTEM] Initializing scraper cluster: 64 virtual worker nodes...
[04:39:01] [RESOLVER] Validating target domain robots.txt sitemaps...
[04:39:01] [PROXY] Rotating exit pool: 12,894 active residential nodes loaded.
ACTIVE NODES: 14,208
INGESTION SPEED: 94,821 p/sec
DEDUPLICATION PLAYGROUND
Shingle size:3-gram
Similarity limit:60%
SHINGLES COUNT A: 12
SHINGLES COUNT B: 11
JACCARD SCORE: 0.0%
ORIGINAL PASSED
02 / DE-DUPLICATION ENGINE

MinHash LSH Deduplicators

For generative models, repeating web page code elements represent overfitting risks. Standard deduplication methods require comparing every page to every other page, which becomes computationally impossible at billion-page scales.

Reworkd resolves this by chunking pages into 3-word shingles, generating integer hash keys, and selecting the minimum hash outputs (MinHash signatures). By partitioning signatures into index bands (Locality-Sensitive Hashing), workers find high Jaccard similarity indices in sub-millisecond rates, stripping repeating boilerplate rows automatically.

  • Locality-Sensitive Hashing
  • Word 3-Gram Shingling
  • MinHash integer signature signatures
  • Dynamic Jaccard parameters
03 / VECTOR ALIGNMENT

Semantic Embedding Pools

Vector databases act as the knowledge retrieval system for Retrieval-Augmented Generation (RAG) models. Reworkd pipelines automatically format scraped text blocks into semantic units using recursive sliding character tokens.

We parse clean Markdown chunks, calculate text embeddings using host models (such as Cohere Embed or OpenAI text-embedding-3), and insert vectors directly into databases.

  • 1536-dim float embeddings
  • Sliding overlap padding
  • Direct vector store sync
  • Hugging Face parquet buffers
SLIDING WINDOW CHUNKERACTIVE PREVIEW
Chunk Size:120 chars
Overlap padding:30 chars
Output token partitions:
CHUNK #1 [OFFSET 0:120]"Large Language Models require high quality pretraining corpora. The quality of crawled text must be pristine to avoid gr"
CHUNK #2 [OFFSET 90:210]"t must be pristine to avoid gradient instabilities and overfitting. Standard HTML extractors generate repeating boilerpl"
CHUNK #3 [OFFSET 180:300]"rs generate repeating boilerplates, template elements, and navigation links. By applying shingle grouping and PII scrubb"
CHUNK #4 [OFFSET 270:346]"hingle grouping and PII scrubbing, Reworkd compiles clean Markdown segments."
Select target destination DB:
from pymilvus import Collection, connections
connections.connect("default", host="api.reworkd.ai", port="19530")
schema = Collection.load("pretraining_corpus")

# Insert payload containing crawled lineage metadata
schema.insert([
  {"name": "vector", "type": "float32", "values": [0.1425, -0.0982, ...]},
  {"name": "source_url", "type": "string", "values": ["github.com/trending"]},
  {"name": "lineage_sha", "type": "string", "values": ["sha_89b2c01"]}
])