qsum

An extractive digest of long text (logs, command output) through a local OpenAI-compatible chat endpoint such as llama.cpp or vLLM. The model is instructed to copy the important lines verbatim and never interpret; qsum then checks that every quoted line really occurs in the input and warns about any that do not. It exists to keep large outputs out of an expensive context window. qsum is only a client: it contains no model and trains nothing. Standard library only.

Install

cp qsum ~/.local/bin/ && chmod +x ~/.local/bin/qsum

Configure

• QSUM_ENDPOINT: first endpoint to try. QSUM_FALLBACKS: comma separated list tried next (default http://127.0.0.1:8080). Each is probed at /v1/models for 3 s.
• QSUM_MODEL: model id (default: first id the endpoint lists). QSUM_KEY: optional bearer token.
• QSUM_VERIFY_LOG: mismatch log (default ~/.cache/qsum/verify.log).
• Flags: --focus "text", --verify (always check; otherwise a random 20% of runs), --chunk N (characters per request, default 9000).

Example

make test 2>&1 | qsum --focus "failing tests" --verify

License

Apache-2.0, see LICENSE and NOTICE. Run the tests with python3 -m pytest.