[Feature] Multi-model index support — separate indexes per model

[rawwerks]

Problem

colgrep currently ties one index to one model per directory. If you switch models via --model, the existing index becomes incompatible and must be rebuilt. This makes it impractical to use multiple models on the same codebase — e.g., LateOn-Code-edge for code search and Reason-ModernColBERT for prose/docs search.

Use case

With the release of Reason-ModernColBERT (congrats on BrowseComp-Plus!), there's a compelling reason to run two models side-by-side:

• Code model for searching code (tree-sitter structured representations)
• Reason model for searching docs, journals, markdown, meeting notes

Today this requires choosing one model or constantly re-indexing when switching.

Proposed solution

Include the model identity in the index directory hash, so different models get separate index directories for the same project.

Current: {project_name}-{hash(path)}
Proposed: {project_name}-{hash(path:model)} (when a non-default model is used)

This is backwards compatible — the default model continues to use the legacy path-only hash, so existing indexes work without migration.

Implementation

I've prototyped this on my fork: rawwerks/next-plaid@feat/multi-model-index

Changes (243 insertions, 59 deletions across 6 files):

• compute_index_dir_name() accepts optional model parameter
• New get_index_dir_for_project_and_model() function
• New IndexBuilder::with_model_identity() constructor
• Search path uses model-aware index resolution
• IndexState gains model_id field

Known gaps in the prototype (would fix before any PR):

1. find_parent_index is not model-aware — could return wrong model's parent index for subdirectory searches
2. model_id is added to IndexState but not yet populated on save
3. status and clear commands don't accept --model flag yet
4. Builder construction pattern is repeated (DRY violation)

Questions for maintainers

Before investing in a full PR, I'd like to know:

1. Is this a direction you'd want to go? Or do you have a different approach in mind for multi-model support?
2. Would you prefer the model identity to live in ProjectMetadata (project.json) rather than IndexState (state.json)?
3. Any concerns about the hash-based separation approach vs. e.g., named profiles?

Happy to clean up and submit a proper PR if there's interest.

[raphaelsty]

Hi @rawwerks, it seems like a good idea to me, you are welcome to open a PR, find_parent_index should be aware of the model of course. clear should clear only the current index belonging to the activated model. So if you are within a directory and you call clear, then you solely clear the index directory which belong to your activated model. I'd rather like the model to be stored within the state.json

Hash-based separation lgtm

This point is for another issue, right now the way colgrep handle text documents is quite basic, maybe we could rely on an external chunking tool for text based documents which is available in rust