PrismRAG — Taxonomy-Controlled Graph RAG Library
Taxonomy-controlled Graph RAG for pgvector, ChromaDB, Pinecone and Weaviate. Explicit rules, Louvain communities and auditable retrieval. Apache-2.0.
Your mapping drives ingest, graph construction, and retrieval—not Microsoft GraphRAG-style document co-occurrence. Free, open-source, runs in your environment.
Interactive demo: github.com/insightitsGit/prismrag/tree/main/examples/demo_app (python demo.py). pip install "prismrag-patch[graph]" (Apache-2.0, v0.2.1). MemoryStore for tests, PostgresStore for production.
What PrismRAG is
PrismRAG is a taxonomy-controlled Graph RAG library (package prismrag-patch 0.2.1, Apache-2.0).
Capabilities: user-defined word-to-category mapping; dual semantic and personal vectors; Louvain communities; BFS graph expansion with semantic re-ranking; MemoryStore and PostgresStore; pgvector, ChromaDB, Pinecone, and Weaviate adapters.
Core features
Tier-1 rules mapping
Auditable word→category assignment with 768-d semantic → 256-d personal projection. Deterministic: same rules produce the same vectors on every run.
Graph RAG search
Community seeding → BFS expansion on your word graph → semantic re-rank. Direct fallback ensures results even when graph traversal finds no path.
Louvain communities
Automatic topic clusters with optional LLM-generated labels for each community centroid. Inspect clusters with rag.list_communities().
Bridge vectors
Synthetic connectors between communities for cross-domain queries — link "medication" ↔ "lab_results" for holistic clinical retrieval.
Append without retrain
Upsert new chunks and merge new mapping rules incrementally. Chunk quality scoring (confidence, separation, coherence) flags low-quality assignments.
Vector DB adapters
Drop-in remap layer for pgvector, ChromaDB, Pinecone, and Weaviate — adds category metadata on insert and search without moving your data.
Use cases by domain
Healthcare (demo)
Categories: medication · lab_results · symptoms. Replace co-occurrence GraphRAG bleed: map insulin/metformin → medication, troponin/HbA1c → lab_results, fever → symptoms. Query “medications for diabetes” returns medication chunks only. Run it: examples/demo_app → python demo.py.
Finance
Categories: risk · valuation · liquidity · regulatory. Enforce that "VaR" and "volatility" always land in the risk category even when embeddings blur lines with "growth." Every retrieval traces to your taxonomy — unlike Microsoft GraphRAG’s corpus graph.
Legal
Categories: contracts · case_law · compliance. Isolated per-matter mappings for regulated document retrieval. Every retrieved chunk links back to the explicit rule that placed it—audit-ready by design.
Enterprise KB
Categories: per-unit mapping JSON. Each business unit submits a mapping JSON; the same SharePoint export yields a completely different knowledge graph per unit—no cross-department bleed.
Public evaluation harness
Type: public parity and evaluation harness. Domains: healthcare, pharmacy, finance. Assertions: ingest completes; at least one community and edge are produced; search returns results; chunk quality is reported.
Integration checks: 13 tests — Package import, no-license operation, ingest, dual vectors, communities, search, category filtering, top-k, per-search latency under three seconds, append, quality reporting, and bridge creation when multiple communities exist.
These are product-authored integration checks using the demo mapping. Runtime varies by environment; no fixed full-pipeline time is claimed.
The public evaluation is a repository test harness, not a neutral cross-vendor benchmark. No comparative accuracy or speed claim is made. Teams should evaluate retrieval quality on their own taxonomy, documents, embedding model, and false-positive budget.
- 13/13 Integration tests — All pass against prismrag-patch==0.2.1 from PyPI (Python 3.12, MemoryStore, offline)
- 0.29 s Demo pipeline — Full ingest → graph → communities → search on healthcare mini-mapping
- Free Apache-2.0 — No license key, no API calls — pip install prismrag-patch and run in your environment
Capabilities
Tier-1 Rules Mapping
Auditable word→category assignment; 768-d semantic → 256-d personal projection. Same rules → same vectors every run.
Graph RAG Search
Community seeding → BFS on your word graph → semantic re-rank. Direct fallback when the graph path runs dry.
Louvain Communities
Automatic topic clusters with optional LLM labels. Bridge vectors link communities for cross-domain queries.
Install
pip install "prismrag-patch[graph]" (Apache-2.0, v0.2.1). No license key. Interactive demo: examples/demo_app.
Pricing
PrismRAG library pricing is $0 forever on PyPI and GitHub. Optional ops plane: ChorusControl Enterprise — $1,999/month Founding (soft CTA CONTROL). See ChorusControl pricing.
Frequently asked questions
What is PrismRAG?
PrismRAG 0.2.1 is an Apache-2.0 taxonomy-controlled Graph RAG Python library. You define categories and word-to-category rules; PrismRAG builds dual vectors, Louvain communities, and an auditable graph retrieval path. Technical guide: https://www.insightits.com/guides/taxonomy-graph-rag.html
Is PrismRAG the same as Microsoft GraphRAG?
No. Microsoft GraphRAG extracts entities and relationships from a corpus. PrismRAG starts from an operator-owned taxonomy and explicit mapping rules. Neither approach is universally better; choose based on whether discovery or policy-owned category boundaries are the main requirement. Comparison: https://www.insightits.com/compare/prismrag-vs-microsoft-graphrag.html
Is PrismRAG free and open source?
Yes. prismrag-patch 0.2.1 is Apache-2.0 with no license key or usage metering — free forever. Soft CTA: TAXONOMY. Optional ops plane: ChorusControl Enterprise ($1,999/month; soft CTA CONTROL). See https://www.insightits.com/products/choruscontrol.html#pricing.
Does PrismRAG send my data to a hosted service?
Not when you run the Python library locally or with your Postgres. Core operation runs in your environment. Your chosen embedding function may call an external model provider, and optional community labeling may call an LLM if you configure it.
Which vector databases does PrismRAG support?
The package provides adapters for pgvector, ChromaDB, Pinecone, and Weaviate. MemoryStore supports tests and notebooks, while PostgresStore provides the production persistence path.
What PrismRAG evidence is published?
The public repository contains 13 demo integration checks and an evaluation harness using healthcare, pharmacy, and finance fixtures. This is product-authored test evidence, not a neutral cross-vendor benchmark. PrismRAG makes no comparative accuracy or speed claim.
What happened to the hosted PrismRAG SaaS?
The Azure SaaS was retired to remove hosting cost. The current product is the Apache-2.0 prismrag-patch Python library, which can run in memory, on your Postgres, or through supported vector-database adapters.
Official package links: PrismRAG source code on GitHub · Install PrismRAG from PyPI · PrismRAG interactive demo
Technical resources: PrismRAG technical guide · PrismRAG comparison · PrismRAG comparison