When assembling context, it’s not enough to just return the right file, the snippet within it needs to be helpful and accurate. We should evaluate + improve how we assemble context for the LLM.

/cc @sourcegraph/search-platform