GraphForge began as a cannabis-specific answer engine and matured into the studio's general-purpose intelligence substrate: a search engine that doesn't just match keywords, but builds a graph of entities and relationships from everything it crawls.

Queries are answered by a hybrid ranker — classic BM25 and TF-IDF signals fused with reciprocal rank fusion — while a schema.org detector lifts structured entities out of pages and links them into a living graph, streamed to the browser over server-sent events as it grows.

BM25 rankingTF-IDFReciprocal rank fusionschema.org extractionKnowledge graphsSSE streamingCrawler engineeringMongoDB

How it works

Crawl. Configurable crawlers ingest target sites, with politeness controls and optional proxy routing.

Extract. A schema.org detector and NLP pass identify entities, types, and relations in each document.

Index. Documents enter dual lexical indexes (BM25, TF-IDF); entities enter the graph store.

Fuse. At query time, reciprocal rank fusion merges the rankers into one result list, enriched with graph context.

Stream. The browser receives the answer and watches the relevant subgraph assemble live over SSE.

FAQ

Why hybrid ranking instead of pure vectors?

Lexical rankers are transparent, cheap, and excellent at exact-match intent; fusion captures much of the benefit of multiple signals without opaque embeddings. The architecture leaves a slot open for a vector ranker as a third fusion input.

What runs on GraphForge today?

CannClicks runs on GraphForge v19 in production, and Risk Radar borrows its ontology-mapping design.

Is it available to license?

Yes — as a hosted engine with your corpus, or as a managed deployment. Contact the studio.