USE CASE · ZERO-TOKEN AI SEARCH

SEARCH YOUR
FILES. SPEND
ZERO TOKENS.

Point XERJ at a folder — source code, a stack of PDFs, a wiki of Markdown — and it indexes everything locally in seconds. Then your team, or your agent, asks in plain English and gets the exact files back. No model runs to search, so nothing leaves the machine and not one token is spent finding the answer. Three real runs below, same one command each time: xerj autoindex <folder>.

SCENARIO 01 · SOURCE CODE

Ask a codebase where something lives.

A new engineer, or a coding agent, needs the line that reads the API key. Instead of grepping guesses or pulling the whole tree into a model's context, ask for it.

the xerj indexer, searching its own source
$ xerj autoindex ./engine/crates/xerj-autoindex/src
phase A sniff + sample → phase B index (20 workers) → finalize
done in 6.7s — 56 files, 112 records, 0 config, 0 secrets indexed
?"read the api key from the authorization header"
esclient.rs … if Some(k) = &self.api_key { r = r.header("Authorization", format!("ApiKey {k}")) …
cli.rs … --api-key <K> Authorization header (or env XERJ_API_KEY) …
0 AI tokens<1 ms3 files ranked
?"sniff the content type instead of the filename extension"
sniff.rs … Content-based format detection. NEVER trusts file extensions. Order: magic bytes … …
0 AI tokens<1 msexact file, first hit
read the folder
~541,000 tokens into a model
ask XERJ
0 tokens

2.1 MB of Rust · reading it all into an agent to find the answer costs ~541k tokens; the query costs none.

SCENARIO 02 · LEGAL DISCOVERY · PDF

Find the case in a stack of filings.

A folder of court PDFs lands on your desk. Somewhere in it is the founder who threatened to walk off with the source code. XERJ reads the PDFs — not their filenames — and ranks the ones that match the accusation.

a folder of court PDFs — read by content, not filename
$ xerj autoindex ./filings # a folder of PDFs
done in 3.6s — 7 PDFs, 18 records, every page's text extracted and indexed
?"a founder threatening to leak source code and customer data unless she is paid"
complaint_meridian_v_okonkwo.pdf … has threatened to leak the Company's proprietary source code and its entire customer database unless the Company buys out her shares at a grossly inflated price …
0 AI tokensPDF text, not filenamerank 1 · 14.1, next filing 5.7
?"someone locked us out of our systems and demanded a ransom to restore access"
complaint_ransom_access.pdf … who deliberately locked the Company out of its own production systems and then demanded payment to restore access … service was unavailable for eleven days …
0 AI tokensa different filing, same folderrank 1 · 14.2, next filing 3.4

Sample filings with fictional parties, for illustration. The PDF text is extracted and indexed locally; no document is uploaded and no model is called to search.

SCENARIO 03 · KNOWLEDGE BASE · MARKDOWN

Answer a project question from the wiki.

Your team's second brain is hundreds of Markdown notes — feature requests, customer threads, release logs. "When did a customer first ask for this, and how long did it take us to ship?" is buried across three of them. XERJ finds them.

a 20-note team wiki, asked by meaning
$ xerj autoindex ./product-kb # a folder of .md notes
done in 5.0s — 20 Markdown notes, 40 records
?"when did a customer first ask for hybrid retrieval?" · semantic recall
customers/acme-freight.md ### 2025-03-12 — Hybrid retrieval ask. Acme raised it on the quarterly call; logged as FR-114 …
feature-requests/hybrid-retrieval.md First requested: 2025-03-12 (FR-114) · requester: Acme Freight …
0 AI tokensmeaning, not keywords
?"which release shipped hybrid retrieval?"
releases/2025-q3.md … Hybrid retrieval (RRF) shipped in v2.4.0 on 2025-09-08
0 AI tokens1 note out of 20
Stitched from notes that share no keyword: Acme Freight first asked on 2025-03-12, shipped in v2.4.0 on 2025-09-08 — about 26 weeks from request to release.

Sample product knowledge base, for illustration. Same one command, a different kind of folder.

HOW IT WORKS

One command, any folder

xerj autoindex <folder> sniffs each file by content, extracts text from code, PDFs, and Markdown, and indexes it — no schema, no pipeline, no config.

No model in the loop

Search is BM25 + semantic retrieval over a local index, the same engine that speaks the _search API. Deterministic, private, and free — zero tokens to find an answer.

Secrets stay out

Hidden files like .env and .ssh are skipped by construction, not by a rule you have to remember. Nothing you index leaves the machine.

TRY IT.

$ curl -fsSL https://xerj.org/get | sh
$ xerj --insecure -d ./data &
$ xerj autoindex ./your/folder
# then ask, in plain English:
$ curl localhost:9200/ax-*/_search -d '{"query":{"match":{"body":"…"}}}'