Retrieve/rerank pipeline with Colbert about colbert

[okhat]

Ah I see. You can always pass --run [name] to the retrieval step so that the output file's name is pre-defined, and then you can use the pre-defined path as input to the rerank step. You can launch both of these from the same command, as a wrapper.

Otherwise, we're actively working on a complete refactor of the codebase that has much better API (including this aspect you asked for) and very accurate yet aggressive compression. It's not perfectly stable yet, but I'm happy to invite you to the private beta if you want to try it out.

[okhat]

Is this for batch retrieval?

Single-query retrieval is already in one script.

[littlewine]

Yes, it has to be batch retrieval, otherwise RAM memory usage is extremely large, so impossible