Implement passage-level retrieval

[mcartright]

Need to add passage-level retrieval. Aim for the coolest thing possible, which, I think, is:

• Allow users to specify how each "retrieval item" is processed. Default implementation is 1 result per candidate (i.e. 1 set of positions per doc), but allow for cooler implementations. Could be a class that gets passed in, or an object that's specified.
• Also allow for varying the number of candidates produced, or varying window size. For example, we may start w/ a size/shift of 200/100, but during execution the shift may increase if we notice that overlapping passages have scores that are too similar.

[mcartright]

Progress (mentally):

• create a "DocumentSlicer" trait that can be set to the views?
• also need a generalization of the "ScoredDocument". "ScoredObject". It would be nice to subclass the trait AND companion object.

[mcartright]

From Laura:

In order to visually compare different relevance models for document retrieval, I want to get a snippet (aka passage) for each document.

This how I did it: 1) get the original query and parameters 2) add passage retrieval stuff to parameters 3) loop over scored documents 3.1) set working set of passage to the document name 3.2) issue query to next layer which performs query parsing, traversing, and retrieval

Problem: Hereby, I am re-computing the same relevance model over and over and over again.

I thought, hey, why don't I save the traversed Node, and then tell the retrieval to go hunting with this node.

Problem: I need to modify the working set, which is baked into the traversed node.

Any ideas?