Tessera DB Tutorial: Build a GraphRAG Retrieval System End to End

A small policy-document retrieval system. Each document references other documents (a "see also" graph). A query against this system returns the most relevant documents using three signals:

We fuse all three with Reciprocal Rank Fusion, so a document strong in any signal rises to the top — no score normalisation needed.

Point Tessera at an embedding provider — a hosted model or a local one. From then on, any document you insert has its text embedded automatically. For quick local testing you can start with a deterministic stand-in embedder and swap in a real model before you care about retrieval quality.

Insert each policy document as an entity with a title and body. The body text triggers auto-embedding on insert, and it's indexed for keyword search at the same time. No separate vector pipeline and no separate search service did any of this.

A "see also" relationship is its own fact connecting a source document to a target document. Adding these lights up the graph dimension automatically: a traversal from any document walks its see-also network. Model the links you know — wire-transfer policy to sanctions policy, sanctions policy to suspicious-activity reporting, and so on.

Take a vague, natural-language question — "what do we need to do for a large transfer to a foreign account?" — and run it through the pipeline. Tessera embeds the question the same way it embedded the documents, scores keyword relevance, walks the see-also graph out from the strongest hits, and fuses the three rankings into one list.

The document about wire-transfer limits ranks first (semantic and keyword hit); the sanctions-screening policy follows because it's graph-proximal to it and shares keywords; related policies surface through their graph links even when their wording doesn't match the question.

The fusion is weighted, and sensible defaults lean on semantics first, then keywords, then graph proximity. Shift the balance to fit your corpus:

Feed the ranked results into your prompt. Because every retrieved fact is traceable to its source through native lineage, the context you build isn't just relevant passages — it's passages your AI can cite. A graph-aware assembly can also include each hit's top neighbours, giving the model contextual stitching pure vector retrieval can't provide.

A working RAG pipeline that combines three retrieval signals over the same store — no separate vector database, no separate search service, no sync between them — and every answer traceable to the facts that produced it. The graph dimension other RAG systems lack is here because Tessera was tracking it anyway.