Technology
How XYRA's immigration intelligence actually works.
One grounded engine runs Ask XYRA, Attorney Intelligence, and Workspace. This page is the real mechanism underneath all three — five stages, each one already shipped, not a roadmap.
Stage 1 of 5
Trusted Sources
XYRA's retrieval index is built from official, publicly documented government sources — the Federal Register's public API, the USCIS Policy Manual, Department of Labor PERM and labor-certification data, SEVP program guidance, CBP guidance, and USCIS Newsroom and Alerts.
Every ingested document is content-hashed and versioned, so an update or a rescission is tracked as a change in the record, not silently overwritten. When a policy is superseded, the system knows — and treats a rescinded document differently from the one that replaced it.
Stage 2 of 5
Grounded Retrieval
When a question comes in, XYRA does not ask a language model to recall an answer from what it learned during training. It searches the indexed source material for what's actually relevant, ranks results by source authority and document status, and only then generates an answer — built from what was retrieved, not from memory.
If the retrieved material doesn't sufficiently cover the question, the system says so, rather than filling the gap with a plausible-sounding guess. This is deliberate: a visible "we don't have enough sourced material for this" is a better answer than a confident, ungrounded one.
Stage 3 of 5
Citation Verification
A generated answer can cite a source. XYRA checks, automatically, whether the citation it produced was actually inside the set of documents retrieved for that question. If a citation references something outside what was actually retrieved, it's rejected — not silently passed through.
This distinction has mattered publicly before: in 2023, a federal court sanctioned attorneys who submitted a legal brief containing citations to cases that did not exist, generated by an AI tool with no citation-checking step. XYRA's verification stage exists specifically to make that failure mode structurally harder — not just less likely, but checked, every time, before a citation reaches the reader.
Stage 4 of 5
Attorney Review
Every AI-generated research answer, case strategy, or draft carries a visible review status. Nothing is filed, sent, or relied upon without a licensed attorney's sign-off — the review requirement is a field on the record itself, not a disclaimer appended after the fact, and it can't be permanently dismissed.
Stage 5 of 5
Grounded Drafting
When XYRA drafts a filing or supporting document, it drafts from a matter's own confirmed, extracted facts — not a generic template with placeholders swapped in. Every claim in the draft traces back to something in the matter's record. The output arrives labeled as a draft requiring attorney review, every time, with no exception.
Key Terms
Four terms worth defining precisely.
RAG (Retrieval-Augmented Generation)
An architecture where a model answers by first retrieving relevant documents from an index, then generating a response grounded in what was retrieved — rather than answering purely from what it learned during training.
Grounded retrieval
Search that returns actual source documents ranked by relevance and authority, forming the only material an answer is allowed to be built from.
Citation verification
An automated check that a cited source was actually present in the retrieved set — catching citations to material that was never retrieved, including fabricated ones.
OCR (Optical Character Recognition)
Extracting machine-readable text from scanned or image-based documents — the first step before a document's content can be classified or have facts extracted from it. See the document review page for XYRA's specific pipeline.
Where This Runs
The same engine, three products.
Ask XYRA
Free, sourced answers for individuals.
Attorney Intelligence
Grounded legal research for immigration attorneys.
Workspace
Evidence, strategy, and drafting for law firms.
For the document-specific pipeline — classification, OCR, and structured fact extraction — see AI Document Review. For when a general-purpose AI model is enough on its own, see Why Not ChatGPT?
See the mechanism on a real matter.
Private Alpha is open to a limited number of immigration firms.