In essence, an inference network is built that is made up of document nodes (from the indexed collection), smoothing parameter nodes, Representation concept nodes, language model nodes, belief nodes, and information (or combination of information) nodes.
- A document node contains a representation of a document's feature vectors such as the term vector for a document.
- A smoothing parameter node allow for user-defined smoothing parameters to enter into the query.
- Language model nodes contain probability distributions for different language models and representations for a document.
- Representation concept nodes represent query language nodes that specify how to combine various weights. Such nodes include the ordered window operations (#n) and boolean ANDs (#band).
- Belief nodes added to the network as needed based on the various scoring functions and query formulations.
- Finally, the information nodes combine the various beliefs from the belief nodes to produce a document's score.
Much more detailed information on the Indri Retrieval Model can be found on [Indri Retrieval Model].