Context: evaluate large repos and multi-repo case

[jtibshirani]

When repos are very large or the context includes multiple distinct repos, it can be harder to retrieve highly relevant results. We should evaluate the multi-repo and large repo case using the "golden queries" approach.

Once we finish an initial "research spike", we'll add a list of tasks.

I suspect that good ranking is going to be critical here, given the larger and noisier result set. Ideas:

• Downrank tests and vendored code
• Fix issue where common terms like "info" or "license" dominate results, due to us not properly implementing BM25

/cc @sourcegraph/search-platform

[keegancsmith]

The sourcegraph repo may be a good "simple" target now for multi-repo. Recently I believe we removed most of the docs directory and it is now in a docs repo. I think similiar things may also happen to our CHANGELOG.