The engines doing the compression are downloaded and run on your machine — nothing goes out to a compression service, and there is no key. That is also the cost, which is why `sift_warmup` exists: the first call pays for loading the models. The honest touch is `sift_extraction`'s diff mode, which shows exactly what was thrown away rather than asking you to trust a rate.
A Python server that sits between raw data and the context window. It routes each input to one of three engines — a heuristic sieve for logs, LLMLingua-2 semantic compression for prose, and a local BGE cross-encoder for ranking — and returns the distilled text with an audit header showing token reduction and latency.
- A local file read through the right pipeline instead of raw, with `.pdf`, `.docx`, `.xlsx`, `.pptx`, `.zip` and `.html` converted to Markdown first — `sift_read_file`
- A noise estimate for a file or a string before you spend context on it, ending in a recommendation for which sifter to use — `sift_analyze_file`, `sift_analyze`
- Repetitive log noise stripped while errors and stack traces survive — `sift_logs`
- Prose and chat history condensed at a rate you choose, with a low-fidelity warning appended when vocabulary overlap falls below the threshold — `sift_chat`
- Long documentation put through two stages: structural sieve first, then semantic compression — `sift_doc`
- OCR and PDF extractions cleaned of metadata debris, optionally with a `--- REMOVED CONTENT ---` section so you can check what was cut — `sift_extraction`
- A list of chunks ranked against a query by a local cross-encoder, returning the top results with confidence scores — `sift_rank`
- Token-saving and latency telemetry, a one-shot project setup that writes IDE hooks, and a model warm-up — `get_sift_stats`, `sift_onboard`, `sift_warmup`
No API key: every model runs locally. `pip install "semantic-sift[neural]"`, then run `semantic-sift`. Reads outside the workspace root are blocked unless `SIFT_ALLOW_GLOBAL_READS=true`. `SIFT_RANK_TOP_N` sets the default result count for `sift_rank` (default `3`), and `SIFT_COMPACTION_FIDELITY_THRESHOLD` (default `0.3`) decides when a compaction is flagged as low fidelity. The package is `semantic-sift` 0.3.8; the `[multi-modal]` extra adds MarkItDown for binary formats.
One command — pip install "semantic-sift[neural]"