Product · Graph

Search that learns the connections — and leads to the goal.

Four graphs built offline from the events AACSearch already collects. The query graph is live today; the product graph, GraphRAG and path-to-goal ranking are in development and marked as such below.

Four graphs

Each graph is built offline from existing data and used online with no model call in the request path.

Query graph

Learns nightly how shoppers move from one query to the next — corrections, narrowings, alternatives — and which transitions end in a click or purchase. Already powers zero-result rescue, “did you mean”, refinements and presearch.

Product graph

Will connect items viewed together, viewed next, substituted and bought together, with hard requires and excludes relations for compatibility.

Knowledge graph (GraphRAG)

Will extract entities from structured fields first and from text only where needed, group them into communities — shopping missions like “first bathroom renovation” — and summarise each one.

Path to goal

Will estimate, for every query cluster, category and item, the probability that a session passing through it ends in a purchase, and rank toward high-value next steps.

How it works with the search engine

Signals in the index

Will write graph rank, purchase value and community membership into each document, so the engine can filter, facet and sort on them.

Graph channel

Will fetch graph neighbours of the top results through the same tenant and permission filters, then fuse them with text results.

Never above the rules

The graph only boosts. Access rules, exclusions and availability stay hard constraints it cannot override.

Available today

The building blocks the graph extends.

Clicked-together recommendations

Session co-clicks already power a recommendation mode.

Personal ranking

Decayed visitor history reshapes search results.

Cited AI answers

RAG over your indexes; GraphRAG will add cross-document context.

Specifications

Query graph
live — rebuilt automatically from search sessions
Product graph (behaviour)
live — co_view, next, complement
Graph database required
no
Language model in the request path
no
Compatibility from catalogue
in development
GraphRAG communities
in development

Before you start

  • Search and click events flowing for at least a few days
  • Basket events for complements

Questions