Audit-ready by default
Every build ships a signed manifest, provenance records, license checks, and benchmark evidence.
Knowledge layer only · no chat UI · no hosted model
RAG datasets and retrieval indexes your team can audit.
Coltex is a production knowledge layer for retrieval-augmented generation. It turns a synthetic document corpus into signed, versioned artifacts—chunks, embeddings, graph edges, metadata, benchmarks, and manifests—so teams can ship RAG systems with traceable data instead of one-off indexing scripts.
10ksmoke documents
≤5%duplicate chunk gate
90%+metadata accuracy target
0third-party docs copied
What Coltex is
Designed to sit underneath your application stack.
Coltex is the data and retrieval layer: document corpus, export pipeline, vector index, graph traversal, benchmarks, and compliance evidence. Bring your own chat interface, model host, API server, and deployment workflow.
Why it is better
Coltex replaces fragile RAG data prep with a repeatable product pipeline.
More than vector search
Chunks, embeddings, metadata filters, graph links, and re-ranking work together as one retrieval layer.
Built for teams
Stable artifact paths and make targets make local validation and cluster builds follow the same workflow.
Artifact set
Versioned outputs with paths developers can wire into pipelines.
Chunks
data/product/chunks/chunks.jsonl
Text segments sized for vector search with typed metadata attached.
Embeddings
data/product/embeddings/embeddings.jsonl
Sentence-transformer vectors for loading into the retrieval index.
Graph edges
data/product/graph/edges.jsonl
Typed relationships such as depends_on and see_also for expansion.
Manifest
data/product/manifest.json
Build metadata and SHA-256 checksums for distribution review.
Retrieval flow
Small, inspectable steps instead of a black box.
1
Encode query embedding
2
Search the vector index
3
Apply metadata filters
4
Expand through graph links
5
Re-rank and assemble context
Quality gates
Validation is part of the build, not a separate spreadsheet.
Product exports run duplication checks, metadata validation, retrieval evaluation, license review, and provenance audits before distribution.
make validate-product
make audit-distribution
make evaluate
Build tiers
Use the same pipeline locally and at cluster scale.
Smoke
Local validation build with 10,000 documents.
make product-premium-smokePremium
Full configured premium dataset export.
make product-premiumHyper
Streaming generation for distributed compute.
make product-hyperQuickstart
Build a dataset, index it, and retrieve context.
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
make product-premium-smoke
make index
python3 -m brain retrieve "What is retrieval-augmented generation for code?" --context