Documentation
Everything Fily does, on one page.
What happens to a file between the moment you upload it and the moment you download it — the formats, the checks, the review, the billing and the API. No login required to read it.
Last updated: August 12, 2026
What actually happens inside
Three stages, the same three for every format. The parts that break translations — tags, structure, terminology — are handled outside the model, not delegated to it.
Step 1
Extract
Your file is parsed into clean, translatable segments. Inline tags, styles, tables, timecodes and document structure are lifted out and held aside, so the model never sees markup it can break.
Step 2
Enforce
A frontier LLM translates each segment with surrounding context, while your glossaries, forbidden terms, do-not-translate lists, numbers and tags are enforced and checked pass after pass.
Step 3
Deliver
You get the same file back, translated and re-importable, with an HTML QA report attached. TM 100% and 101% matches are copied byte for byte and never touched by the model.
Supported formats
Each format has its own pipeline — a Trados file and a scanned PDF do not go through the same code. Follow any link for what that pipeline does and where that format usually breaks.
What runs on every job
None of this is an add-on tier or a checkbox you have to remember. It is what the pipeline does by default, on every file.
A 12-step QA pass
Glossary enforcement, forbidden-term replacement, number and date checks, consistency, tag-count repair and seven more passes — every one of them auditable in the QA report attached to your delivery.
Your glossaries and style guides, enforced
Upload TMX, XLSX or DOCX glossaries and define style guides — do-not-translate lists, forbidden terms, tone instructions. They cascade organization → client → sub-client → job, and the most specific rule wins.
Translation memory that stays intact
Segments already marked 100% or 101% are never sent to the model and never billed. Only new and fuzzy segments are translated, so a pre-translated file comes back with its match status untouched.
Tag repair and round-trip verification
Tags and placeholders are masked before translation, reinserted in their original positions, then counted and repaired — so the file re-imports into your CAT tool instead of erroring on return.
Cross-file consistency across a batch
In a fifty-file batch the same source sentence used to come back three different ways. A tag-safe consistency pass runs over the whole batch, flags every divergent span, and unifies it in one click.
Back-translation for regulated content
Each segment is translated back into the source language by an independent pass and compared for semantic drift, so a reviewer who does not read the target language can still sign off.
Review & Human QA
The workflow is invisible by default, not opaque. Everything the pipeline decided is on screen if you want to look at it — and editable if you want to change it.
The browser review editor
Every job can be opened segment by segment, source beside target, with the QA findings and the back-translation in the same view. Edits go into the delivered file — the deliverable is rebuilt from your corrections, not patched on top of them.
Quality scoring in the MQM tradition
Defects are classified on six dimensions — accuracy, terminology, fluency, markup, consistency, locale — across three severities. The score is normalized per 1,000 words, so a long job cannot dilute its own errors.
Human QA, one toggle away
Flip Human QA on any job and a professional linguist works through the output before delivery, starting from the QA report and the back-translation. Quoted upfront and tracked in your dashboard.
Reviewer accounts for your client's editors
Your client's own reviewers get accounts that see only the jobs assigned to them — never your other clients, your volumes or your configuration. Long jobs can be split into stretches across several reviewers.
Delivery and the client portal
Your clients get a branded space where they both start the work and receive it — so the job never has to live in an email thread.
- Deliverables are downloaded from the authenticated portal — never sent as email attachments.
- Batches come back as a ZIP, file by file, with the QA report per file.
- Your clients can place requests in the portal themselves: files, languages, service, deadline and reference files that are never counted, charged or translated.
- Requests land in a queue on your side and pre-fill the job form — nothing sits unread in an inbox.
Billing
Everything runs off one balance, whatever the job was. Words are billed per word, OCR and PDF work per page, audio per minute — all at the rates of your plan, all drawn from the same pool.
- Every job is quoted before it runs — you approve the number, and it is the number you pay.
- TM 100% and 101% matches are excluded from the count entirely.
- Reference files attached to a request are never counted or charged.
- Subscribe for a monthly allowance, or top up with prepaid packs at any time.
API & webhooks
On Pro and above, submit jobs, poll their status and pull deliverables from a REST API, or let signed webhooks tell your system the moment a job changes state — no polling loop of your own.
Tokens and webhook endpoints are managed from your account. The reference is published and maintained; ask your Account Manager for a key and point your system at it.
Request API accessSecurity & residency
TLS in transit and encryption at rest, per-organization isolation, single sign-on and two-factor authentication, configurable retention, and portal-only delivery. Your content is never used to train models — ours or anyone else's.
If the requirement is jurisdictional or architectural, the same pipeline runs pinned to a region, inside your own cloud account, or fully on-premise on open-weight models.
Frequently asked
The fastest way to read the rest of this is to send us a file.
Upload something real — the file your current tool chokes on — and read the QA report that comes back with it.