Compatibility

locus index walks a folder and, per file, checks two things: is the extension in the indexed set, and is the file at or under a size cap (default 25 MB, LOCUS_MAX_FILE_MB to change it). A file that fails either check is skipped entirely — never partially indexed. The table below comes directly from the parser's own format list, not a general claim about what "text extraction" might handle.

A handful of formats the parser supports (.htm, .xml, .sh, .sql) aren't walked by default — add them with LOCUS_EXTENSIONS (comma-separated, with or without the leading dot) if you want them included. PDFs are extracted as their text layer only: a scanned or image-only PDF with no embedded text returns empty content, since Locus does text extraction, not OCR. Symlinks are never followed, hidden files/folders are skipped, and common noise directories (.git, node_modules, .venv, dist, build, __pycache__, and similar) are always excluded — the same rules whether you're on the free local engine or the hosted trial, since it's one indexing codebase either way.

The free local mode is optimized for English; managed mode uses a multilingual model. We have not yet measured search quality in other languages. Non-English files are still parsed and indexed the same way; what changes is how well search finds them. The local default is an English embedding model, so expect non-English files to be found less reliably.

Once a file is indexed, Locus hands back the same thing whichever client asks: chunk text plus a file path, from locus_search and locus_read_file. The client never parses the original file itself — Locus already did that at index time — so a client's ability to use a result has nothing to do with whether that result came from a .md file or a .pdf. A client that can talk to Locus at all can query any indexed format equally; there is no per-format client restriction anywhere in this architecture.

What does vary per client is how far we have tested it ourselves. Each client carries one of four labels, the same ones used on the full connection guide, and each label rests on a test record we keep:

In short: format decides if a file is in the index at all; client decides how you ask questions about what's already in there. A Tested client asking about a .docx file and a Documented client asking about a .py file go through the exact same retrieval path once indexing is done — neither axis constrains the other. Full per-client setup steps live on /docs/connect; what's free versus trial versus paid to run any of this lives on /free, and how Locus fits alongside the note app, cloud storage, and AI assistant you already use is on /complements.