llm-id-compressor

Shrinks LLM prompt/response token counts when you're sending a list of UUID-keyed items (transcript lines, database rows, log entries) to a model. Swaps each UUID for a short numeric ID before the request, and restores the original UUIDs from the model's response afterward.

Install

pip install llm-id-compressor

Usage

from llm_id_compressor import create_id_mapping, replace_uuids_with_nums, restore_uuids_in_response

items = [{"id": "550e8400-e29b-41d4-a716-446655440000", "text": "..."}, ...]

mapping = create_id_mapping(items)
compact_items = replace_uuids_with_nums(items, mapping["uuid_to_num"])

# send compact_items to the LLM, get back a list of {"id": "<num>", ...} lines
response_lines = call_llm(compact_items)

restored = restore_uuids_in_response(response_lines, mapping["num_to_uuid"])

replace_uuids_with_nums expects each item to have exactly id and text keys - any other fields are dropped, and an item missing text raises a KeyError. restore_uuids_in_response has no such restriction: it preserves every field on each response line, only rewriting id.

restore_uuids_in_response silently drops any line whose numeric id wasn't in the original mapping (a hallucinated id) rather than raising - callers that insert results by a global primary key would otherwise have one bad line abort the whole batch.

Why this exists

Sending a list of UUID-keyed items to an LLM burns tokens twice over: once in the prompt, once in the response, on 36-character identifiers the model never actually reasons about. There's no other package on PyPI doing this specific swap - it's a five-minute fix once you know to look for it, but easy to miss.

Contributing

Issues and PRs welcome - see CONTRIBUTING.md.

License

MIT