iternal-technologies-partners/blockify-agentic-data-optimization

Transform messy enterprise content into compact, validated knowledge units optimized for AI

Patented data ingestion, distillation, and governance pipeline. IdeaBlocks replace naive chunking with structured, deduplicated, LLM-ready knowledge.

78X Aggregate Performance  ·  2.29X Vector Search Accuracy  ·  29.93X Distillation  ·  3.09X Token Efficiency  ·  40X Size Reduction

 

 

 

---

What is Blockify?

Traditional Retrieval-Augmented Generation (RAG) pipelines split documents into fixed-size chunks, then hope that vector similarity will surface the right context. It rarely does. Chunks break mid-sentence, duplicate content inflates token bills, and hallucinations slip through because the LLM is reasoning over fragments rather than facts.

Blockify replaces naive chunking with a patented ingestion and distillation pipeline that transforms raw enterprise content into IdeaBlocks — structured, semantically complete XML knowledge units. Every IdeaBlock carries its own question, trusted answer, tags, entities, and keywords. Similar blocks are deduplicated and merged so the knowledge base stays compact, coherent, and governable.

[!TIP] An IdeaBlock is the smallest unit of curated knowledge: a self-contained XML unit with a name, critical_question, trusted_answer, tags, entity, and keywords. Unlike fixed-size chunks, IdeaBlocks preserve full semantic coherence.

---

How It Works

1. Ingest — Documents (SharePoint, Confluence, Git, local docs) are parsed and transformed into structured IdeaBlocks via the Blockify API
2. Distill — Similar blocks are clustered with embeddings + LSH, then merged by the distill LLM to eliminate duplicates while preserving distinct facts
3. Retrieve — Optimized blocks are stored in vector databases (ChromaDB, Pinecone, Cloudflare Vectorize, Neo4j) for high-accuracy RAG retrieval

---

Features

---

Repository Structure

This repository contains two deployable components plus comprehensive technical documentation:

---

Quick Start

[!NOTE] New to Blockify? Sign up for free to get $1,000 in API credits, then pick the path below that fits your workflow.

Path 1 — Test the API in 30 seconds

curl --location 'https://api.blockify.ai/v1/chat/completions' \
  --header 'Authorization: Bearer YOUR_API_KEY' \
  --header 'Content-Type: application/json' \
  --data '{
    "model": "ingest",
    "messages": [{"role": "user", "content": "Your text to process here"}],
    "max_tokens": 8000,
    "temperature": 0.5
}'

Path 2 — Claude Code Skill (5 minutes)

git clone https://github.com/iternal-technologies-partners/blockify-agentic-data-optimization.git
cd blockify-agentic-data-optimization/blockify-skill-for-claude-code/skills/blockify-integration

pip install -r requirements.txt
python3 scripts/setup_check.py

# Ingest your docs, distill, and run semantic search
python3 scripts/run_full_pipeline.py --source ./my-docs

See blockify-skill-for-claude-code/ for full skill documentation.

Path 3 — Distillation Service (Docker)

cd blockify-distillation-service
cp .env.example .env
# Edit .env with your BLOCKIFY_API_KEY and OPENAI_API_KEY

docker-compose up -d

# Verify
curl http://localhost:8315/healthz

---

Enterprise Edition

[!IMPORTANT] The open-source components in this repository are a fully capable starting point. For production workloads at enterprise scale, Blockify Enterprise provides productionized containers with significantly expanded capabilities.

What Enterprise Adds on Top of Open Source

Who is Enterprise for? Teams ingesting millions of documents, running regulated workloads, needing automated refresh pipelines, or deploying in air-gapped environments.

 

---

API Models

The Blockify API exposes three models via an OpenAI-compatible chat completions endpoint:

See BLOCKIFY-API-REFERENCE.md for full endpoint documentation.

---

The IdeaBlock Format

Every IdeaBlock is a self-contained XML knowledge unit:

  Blockify Overview
  What is Blockify?
  Blockify is an agentic data optimization pipeline that converts unstructured enterprise content into compact, deduplicated XML IdeaBlocks to improve retrieval accuracy and reduce token usage in RAG and LLM workflows.
  RAG, DATA_OPTIMIZATION, KNOWLEDGE_MANAGEMENT
  
    BLOCKIFY
    PRODUCT
  
  blockify, ideablock, RAG, distillation, deduplication, enterprise AI

---

Performance

[!TIP] Prove the numbers on your own data: run the built-in benchmark suite from the Claude Code skill — python3 scripts/run_benchmark.py --company "Your Company" — and get an HTML report comparing IdeaBlocks vs. traditional chunking.

---

Documentation

---

Integrations

Blockify sits between your document source and your retrieval / storage layer. It plugs into every major RAG framework, vector database, data platform, and workflow engine. Each guide below covers the problem Blockify solves on that stack, an architecture diagram, quick-start code, advanced patterns, and a side-by-side comparison with the platform’s default behavior.

RAG Frameworks

Knowledge & Workflow

Vector & Search Databases

Data Platform & Observability

Document Parsing

[!TIP] See the full integrations index for pattern references. Don’t see your stack? Blockify exposes an OpenAI-compatible API — if your platform can POST HTTP, it can use Blockify.

Generic Pattern

Documents -> Parser -> Blockify (Ingest + Distill) -> Embeddings -> Vector DB -> LLM / Agent

Claude Code — Install the skill (see Path 2) to let Claude Code ingest project documentation into local ChromaDB and perform high-accuracy semantic retrieval during development work.

Chatbot & Customer Support — Blockify-processed knowledge reduces hallucination risk and improves answer quality in production chatbots. See OPENCLAW-RAG-INTEGRATION.md for a Cloudflare Workers example.

---

Frequently Asked Questions

What is Blockify and how is it different from naive RAG chunking?

Blockify is a patented ingestion and distillation pipeline that replaces fixed-size text chunking with IdeaBlocks — structured XML knowledge units containing a name, critical question, trusted answer, tags, entity, and keywords. Unlike RecursiveCharacterTextSplitter (LangChain) or SentenceSplitter (LlamaIndex), Blockify is semantically aware, deduplicates across the corpus, and produces ~2.5% of the original dataset size while preserving 99%+ information fidelity.

How does Blockify improve vector search accuracy?

On real enterprise corpora, Blockify delivers 2.29X improvement in vector search precision (average-distance-to-best-match from 0.3624 to 0.1585). Because duplicates are collapsed before vectorization, the top-K neighbors are semantically distinct rather than near-duplicates of the same boilerplate.

Does Blockify work with my existing vector database?

Yes. Blockify is embedding-model and vector-database agnostic. There are dedicated integration guides for Milvus, Zilliz Cloud, Elastic, Supabase, and Cloudflare Vectorize. Pinecone, ChromaDB, Qdrant, Weaviate, and pgvector work through the LangChain and LlamaIndex adapters.

Does Blockify replace LlamaIndex or LangChain?

No — Blockify composes with LlamaIndex and LangChain. It replaces the chunking stage (NodeParser / TextSplitter) with a higher-quality transformer that produces IdeaBlock-backed nodes or documents. See Blockify + LlamaIndex and Blockify + LangChain.

How does Blockify reduce LLM token costs?

Blockify delivers 3.09X token efficiency — ~98 tokens per retrieved block vs. ~303 per traditional chunk. On 1B queries/year this translates to ~$738,000 in token cost savings. Cost reduction comes from two compounding sources: (1) 40X fewer embeddings to generate and store, and (2) denser retrieved context means fewer tokens per LLM call.

What’s the difference between Blockify Ingest and Blockify Distill?

Ingest converts raw text into draft IdeaBlocks. Distill clusters similar IdeaBlocks (embeddings + LSH + LLM synthesis) and merges them into canonical blocks, eliminating duplicates. A typical pipeline runs Ingest per-document then Distill across the corpus. Both are exposed as models via the same OpenAI-compatible API.

Can Blockify run air-gapped / offline?

Yes. The open-source distillation service runs fully offline with a local LLM runtime (VLLM, NVIDIA NIM, Intel OpenVino). Blockify Enterprise ships air-gapped deployment (AirgapAI) with hardened containers for classified environments.

Is Blockify open source?

The Claude Code skill, distillation microservice, Helm chart, benchmark suite, and 12 integration guides in this repository are open source under the Blockify EULA. Blockify Enterprise adds hardened containers, scheduled connectors, RBAC, governance dashboards, and professional services — see Enterprise Edition.

How do I migrate from naive chunking to Blockify?

Three paths: (1) run the Claude Code skill against your docs for immediate local results, (2) add a single HTTP call to api.blockify.ai/v1/chat/completions in your existing pipeline, or (3) self-host the distillation service via Docker or Helm. The Getting Started Guide walks through all three.

What document formats does Blockify support?

Via Unstructured.io or native parsers: DOCX, PDF, PPTX, PNG/JPG (OCR), Markdown, HTML, email. See Blockify + Unstructured.io for the recommended parsing pipeline.

---

Contributing

Contributions are welcome. See CONTRIBUTING.md for development setup, code style, and the pull request process.

[!NOTE] This project is governed by the Blockify Community License. By contributing, you agree that your contributions will be subject to these terms, including the Contribution License (Section 2.4).

Security issues should be reported privately — see SECURITY.md.

---

Support & Community

---

---

License

This project is licensed under the Blockify Community License. Free for developers, researchers, and companies under $1M annual revenue. Organizations exceeding $1M annual revenue require an enterprise license — contact [email protected]. See ENTERPRISE.md for details on enterprise capabilities.

Blockify, IdeaBlock, and AirgapAI are trademarks of Iternal Technologies, Inc.