Document Analysis
Let visitors upload a PDF, DOCX, XLSX, CSV, or TXT in chat — instant translation quote, resume screening, summary, document checklist, or your own template.
Beta. This feature is still being tested, and how it behaves may change. The tile and the side panel show a Beta label. It works and is safe to use, but some documents may still give imperfect results — use Test with my file (see Check that it works) to see how a template handles a real document before you rely on it.
Audience: Public — for customers configuring analysis templates and visitors using them.
What it does
Document Analysis lets your visitors upload a document directly in the chat widget — a PDF, DOCX, XLSX, XLS, CSV, or TXT — and get a structured answer back in seconds. It's different from the knowledge base, where you add your own material for the agent to answer from. Document Analysis processes one visitor's document per chat turn and returns a result tailored to them — then forgets it.
Use it when your customers regularly send you documents to estimate, screen, summarize, or extract from:
- Translation agency — visitor uploads a source document → instant word-count-based quote in your currency.
- HR / recruitment — visitor uploads a resume → instant match score against the job description you configured.
- Consultants / legal — visitor uploads a long doc → a TLDR with the bullet count and tone you chose.
- Legal intake / immigration / accountants / mortgage / admissions — visitor uploads their document set → the agent checks it against your required-items list and reports what's present, what's missing, and whether the package is complete.
- Something else — write your own free-form prompt (Custom template) that the AI runs on an uploaded document when the visitor's request fits it.
You set up analysis templates in the dashboard; the AI picks the right one based on the document and the visitor's question, runs it, and replies in the same chat.
How visitors use it
- A paperclip icon appears in the chat input. It shows up only once your plan includes Document Analysis and you've created at least one active template, and it is hidden while this month's upload allowance is used up.
- The visitor clicks it and picks a file. The file picker offers the supported types, and the chat checks the file's type and size before sending it.
- A chip with the filename appears in the chat input.
- They type their question (e.g. "How much to translate this to French?") and send.
- The filename shows inside their message bubble; the agent reads the document and replies with the structured answer.
- Once the reply is sent, the service keeps no copy of the document.
If an upload fails, the visitor gets a short explanation instead — see What visitors see when an upload fails.
Structured cards
For six of the built-in kinds, the widget renders a structured card below the AI's written reply — the reply stays conversational, and the card gives the visitor the hard result at a glance:
- Translation Quote → a price card: the total and currency, the source → target language pair, the word count, and the turnaround.
- Print Quote, Editing Quote, Certified Translation Quote and Accounting Estimate → a service-quote card: the same idea in one shared layout — the total, the currency, what was counted, and the turnaround.
- Document Checklist → a checklist card: a ✓ or ✗ against each required item, and the overall complete state.
The figure on the card is computed by the server from your rates and the parser's count, not by the AI. The AI passes on the choices the visitor states, such as the target language or the number of copies, and is asked to mention the figure briefly rather than repeat the breakdown. The card's figure comes from that calculation, not from the wording of the reply.
Resume Screening, Document Summary, and Custom deliberately stay as plain prose — no card. That's a soft-framing choice: a candidate, for instance, shouldn't be handed a blunt "FAIL" card, so screening results reach the visitor as the assistant's prose rather than as a verdict card.
Supported file types
| Format | Local parser | Premium parser (plans with native reading of complex PDFs) |
|---|---|---|
.pdf |
Text-layer extraction | Gemini native PDF — tables, scans, charts, multi-column |
.docx |
Word text extraction | — (local is already excellent for Word) |
.xlsx, .xls, .csv |
Spreadsheet → text/tables | — (local is native for spreadsheets) |
.txt |
Plain text | — |
How the parser tier is chosen automatically. A document is first read with the fast local parser. On a plan with premium parsing (see Plan and limits), if a PDF comes back low-confidence — the classic case is an image-only scanned PDF where the text layer is empty, or a dense table that didn't extract cleanly — the service automatically sends that PDF to the Gemini native parser and cross-checks the two results before answering. It skips this step for a PDF over the plan's page budget for premium reading, a PDF whose page count is uncertain, or one whose premium read would cost more than a per-document budget; the local result is used then. Clean, born-digital documents stay on the fast local parser, so premium parsing is used only where the local read came back low-confidence. On plans without premium parsing, documents stay on the local parser.
Word, Excel, CSV, and TXT files are read by the local parser alone: the premium parser reads PDFs only, because a scanned PDF is the case where the text layer is empty and a vision model is needed to read the page. These formats carry their text as text. The local parser still reports when it is unsure — a Word file it could pull no words from, a spreadsheet larger than 50,000 rows in a sheet, which it reads up to that point.
What you need
- A plan that includes Document Analysis. Plan and limits states which plans include it, with the upload limits, file-size caps, parser tier and how many active templates you can have on each.
- At least one active template. Setup happens in the dashboard; you need no account with another service and no key.
How to connect
Setting it up
The panel shows two steps: Configure a template (steps 1–4 and 6 below), then Verify it on a real file (step 5).
- In Dashboard → Integrations, open the Document Analysis tile; its side panel opens. If the tile shows Available in … and higher instead, your plan doesn't include the feature yet — the panel shows which plan unlocks it, with an Upgrade to … button.
- Click Add template and pick a kind from the grouped tiles under Template type — see The built-in templates below for the full list (Custom is only selectable on plans where the table under Plan and limits ticks “Your own template wording”; on the others it carries a plan lock).
- Give it a Display name and fill in the per-kind settings (rate, job description, bullet count, required items, prompt, etc.). The editor smooths the fiddly parts for you:
- Must-have skills and Required items are entered as chips — type one and press Enter or a comma.
- Rates for specific language pairs are rows: pick From and To in a searchable list of language names and type that pair's rate — no raw language codes.
- Custom prompts have insert buttons for
{{wordCount}},{{language}},{{pageCount}}, and{{filename}}, plus a warning if you type a placeholder other than these four. - A live price example sits under the Translation Quote params and recomputes shortly after you stop typing (e.g. "Example: 500-word EN → DE document = $60"), so you can see the effect of a rate change before saving. It is worked out on our server without reading a file, so it does not use up your daily Test with my file allowance. If it ever shows "Live example paused — too many requests in a short time", only the example is affected, not your template; it updates again with your next change.
- Toggle Active to enable it. Primary beside it marks the template as primary in your Templates list. It is a label only: it does not change which template the assistant uses.
- Optionally hit Test with my file to see what the template produces for a real file before you save — see Check that it works.
- Click Create template (Save changes when you edit an existing one) — the paperclip icon appears in your widget the next time a page with your chat loads (reload a page you already have open). Cancel, or the cross at the top of the editor, closes the editor; if you have changed something, it asks Discard your changes? first. Cancel and the save button sit at the start of the editor, under its title.
You can have as many active templates as your plan allows; inactive templates do not count toward the limit (see Plan and limits). Each active template becomes its own tool the AI can choose from, by its display name and settings. Premium PDF parsing needs no switch — on plans with native reading of complex PDFs it's used automatically for a PDF the local parser read with low confidence (see Supported file types above).
The template editor is roomy: on a wide screen the side panel opens wide, and Compact and Wide narrow and widen it without losing what you have typed. If you try to close the panel with unsaved changes to a template, it asks Close without saving?
Settings
The built-in templates
You set up analysis templates in the dashboard. The Add template picker offers eight built-in kinds in two groups, plus a free-form Custom kind on plans with your own template wording:
Price from a file — the visitor uploads something and gets a quote:
- Translation Quote — everyday text translation, priced per word.
- Print Quote — print shops and copy centres, priced per page; scans are fine.
- Editing Quote — proofreading and editing, priced per word.
- Certified Translation Quote — sworn or notarized translation and apostille, priced per page; scans are fine.
- Accounting Estimate — bookkeeping firms, a starting estimate by document volume.
Understand a document — the visitor uploads something and gets an answer about it:
- Resume Screening — match a candidate's resume against your job description.
- Document Summary — a short summary of an uploaded document, made from its first 8,000 characters.
- Document Checklist — check the upload against your required-items list.
Your own — Custom, where you write the prompt yourself (on plans with your own template wording).
The tiles that suit your line of business carry a Recommended badge (or Fits and your site's domain) and the others are dimmed rather than hidden, so you can still pick them. A kind your plan doesn't include shows a lock instead.
You can create multiple templates per kind (e.g. two Translation Quote templates with different rates), and the AI picks the one whose display name and settings best fit the visitor's question. In the Templates list, Set primary marks one template with a primary badge. Because the assistant chooses by display name and settings, give each template a name that says what it is for.
Translation Quote
Visitor uploads a document → the agent returns a price for translating it.
You configure:
- Rate per word (e.g.
0.12) - Currency — USD, EUR, GBP, CAD, or AUD
- Minimum quote (e.g.
50— applied whenwords × ratecomes in under it) - Turnaround in business days (e.g.
3) - Rates for specific language pairs (optional) — set a different price per word for particular pairs that cost you more (or less) than your base rate. This section is shown right under the base rate — pick source and target from a searchable language list (you see the language name; the code is stored for you, so a mistyped code like
jpis impossible) and type that pair's own rate. Direction matters —en → jaandja → enare separate rows. A pair that isn't in the list uses your base rate. Examples:English → Japanese=$0.22/word (scarce talent pool)English → Arabic=$0.24/word (right-to-left typesetting overhead)English → Spanishnot in the list → your base rate applies
How the number is produced: total = max(word_count × rate, minimum_quote), where rate is the specific pair's rate if you set one, otherwise your base rate. It's pure local math — no extra AI cost beyond the chat reply itself. The word count is the parser's count, and a word count the AI supplies is ignored, so the price follows the document's real length rather than an AI estimate. A scan or photo PDF with no text layer has no words to count, so the agent asks the visitor for a digital file (DOCX, TXT or a digital PDF) instead of quoting. The source language comes from the parser's detection (unless the visitor overrides it); the target language comes from the visitor (the AI asks if it isn't stated).
Example reply:
"Your document is 1,543 words, detected as English. You asked for Spanish translation. At $0.12 per word that's $185.16 USD, and turnaround is 3 business days. Want me to connect you with a project manager?"
The other quote kinds
Print Quote, Editing Quote, Certified Translation Quote and Accounting Estimate work the same way as Translation Quote — you set the rates, the server does the arithmetic on the parser's own count, and the visitor gets a card. What differs is what gets counted and which extra settings you get:
| Kind | Counted in | You also set | How the total is worked out |
|---|---|---|---|
| Print Quote | pages | an optional colour rate per page, and a minimum order | pages × rate × copies, floored at your minimum. The colour rate is used only if the visitor asks for colour and you configured one. PDF only — for another format the agent asks for a PDF. Scans are fine — page count survives a PDF with no text layer. |
| Editing Quote | words | named service rates (e.g. "Proofreading" at a lower per-word rate than "Copywriting"), a minimum quote, and turnaround | words × the matching service rate, or your base rate when the visitor doesn't name one, floored at your minimum. Digital text only — a scan has no words to count, so the agent asks for a digital file instead. |
| Certified Translation Quote | pages | a certification fee per copy, a minimum quote, and turnaround | pages × rate per page, plus copies × certification fee, floored at your minimum. PDF only; scans are the normal input here and work fine. |
| Accounting Estimate | pages | a base fee it starts from, an optional per-page increment, and a period label (default "per month") | base fee + pages × increment. Deliberately a from figure, not a firm quote — the card is labelled as an estimate, and pages only ever add to the base. |
Resume Screening
Visitor uploads a resume → the agent scores it against your job description.
You configure:
- Job description — free text; the AI uses it as the scoring baseline.
- Minimum years of experience (e.g.
3) - Must-have skills — entered as chips (type a skill and press Enter or a comma, e.g.
TypeScript,PostgreSQL,AWS)
How the verdict is produced: the AI reads the resume and extracts the candidate's observed skills, years of experience, and an overall score. The engine then compares those deterministically against your bar and returns a verdict of pass, borderline, or fail, along with which required skills matched, which are missing, and whether the years requirement was met. The AI is instructed to fill in only what's literally on the resume and not to invent skills or experience.
Document Summary
Visitor uploads a doc → the agent returns a TLDR.
You configure:
- Max bullet points (e.g.
5— the engine trims to this cap if the AI returns more). - Tone —
professional,casual,technical, orexecutive. - Include extracted action items — when on, adds a separate "next steps" list.
If a visitor uploads a document that doesn't fit your templates, the assistant is told to fall back to your Document Summary template when you have one active. Without one, it still has the document's text in front of it and answers from it — often with a summary — rather than leaving them stuck. That free-form answer is the assistant reading the document, not a template running: no template result is recorded for it. (Uploads only work once you've created at least one active template — see How visitors use it above.)
Custom
A free-form prompt with placeholders — use it when the eight built-in kinds don't fit your case. You write the instruction; the AI runs it against an uploaded document when the visitor's request matches it, and returns your structured result. The Custom kind is locked on plans without your own template wording and unlocks on the others (see Plan and limits).
You write a prompt like:
"Extract from this invoice: vendor name, invoice number, total amount, due date. The document is {{wordCount}} words, language {{language}}, {{pageCount}} pages, filename {{filename}}. Return as JSON."
Placeholders the engine fills in at runtime:
{{wordCount}}— words the parser extracted{{language}}— detected ISO-639 code (en, es, ja, …){{pageCount}}— pages (PDF) / sheets (XLSX) / 1 (text){{filename}}— the original uploaded filename
Those are the four placeholders the engine substitutes. Insert buttons in the editor drop each placeholder in at your cursor, and the editor warns you if you type a {{token}} other than these four. The parsed document — up to its first 8,000 characters — is given to the AI alongside your prompt, so it can quote and reason over the actual content, not just the stats — there's no separate "sample text" placeholder because that text is already there.
Document Checklist
Visitor uploads a document set → the agent checks it against a list you defined and reports what's present and what's missing.
You configure:
- Required items — a list of short labels for the documents (or sections) a complete package must contain, entered as chips (type an item and press Enter or a comma, e.g.
Passport,Proof of address,Bank statement — last 3 months). You can list 1 to 20 items.
How the verdict is produced: the AI reads the uploaded document — its first 8,000 characters — and judges, for each required item, whether it's present in the text. The engine then cross-checks those judgements against your list deterministically — your list is the source of truth, the AI only decides presence — and returns a found / missing status per item plus an overall complete / not complete. It won't invent items you didn't ask for.
Typical uses: legal intake, immigration services, accountants collecting tax documents, mortgage brokers, and university admissions — cases where the question is "is the document package complete?".
Check that it works
Before you save a template, open Test with my file in the editor, choose a real document and click Run test (for a Translation Quote, pick the language to translate into first). You see what the template produces for that file — the parser's figures (words, pages, language, which parser ran) and the structured result. Test runs don't use your monthly upload allowance and aren't written to your analytics; they're limited to roughly 20 tests a day, with a small separate premium-parse test budget of about 3 a day, after which tests use the standard parser.
After saving, open (or reload) a page with your website's chat: the paperclip appears in the message box. Upload a document and ask your question.
Watching how it performs
The Document Analysis tab in Website Analytics (which plans have it) shows how well this is working: how many documents were analyzed, the mix by file type / template / parser tier, parsing success and rejection rates, a confidence distribution, per-template outcomes (quote totals, screening verdicts, summary counts, and — for Document Checklist — average items-found and complete rate with the top missing items), thumbs-up/down feedback on document answers, and a paginated list of recent uploads with a "needs review only" filter for the low-confidence ones.
If it isn't working
"The paperclip icon doesn't show in my chat"
- Confirm your plan includes Document Analysis (see Plan and limits).
- Confirm you've created at least one active template in the dashboard.
- Check whether this month's upload allowance is used up — the paperclip is hidden until it resets. The Plan Usage card in Website Analytics shows it.
- Reload the page hosting your widget — the widget reads its settings when the page loads.
"A visitor got a 'document too complex' or low-quality result on a plan without native PDF reading"
- Some PDFs (image-only scans, very complex tables) exceed what the local parser can read well. Move to a plan with native reading of complex PDFs to unlock automatic Gemini-native PDF parsing for those documents, within a per-document budget.
"The quote came back at the minimum for a tiny document"
- That's the minimum-quote floor working as designed. If
words × ratelands below your minimum, the minimum wins. Lower the minimum in the template editor if you want to allow smaller jobs.
"An old Excel file returned garbled text"
- Try saving it as
.xlsxrather than the legacy.xlsformat — spreadsheet support is most reliable on.xlsx.
"I hit my monthly upload cap"
- While the cap is reached, the paperclip is hidden from your chat. The cap resets on the same day of the month as your message quota. If you consistently need more, upgrade the plan — a higher one raises it (see Plan and limits). See Billing & plans.
What visitors see when an upload fails
Two checks happen in the chat before the file is sent: a file over your plan's size limit, or of a type that isn't supported, is turned away there with a short message naming the problem. Past that point, the replies are not canned sentences: the agent is handed a short, visitor-safe phrase describing the problem and writes the reply itself, in the visitor's language and in your persona's voice, so the wording varies.
- File too large → the agent says the file is too large.
- PDF too long → the agent says the document has too many pages; the page limit is stated below this list.
- Wrong file type → the file picker offers the supported types, and the chat turns away a file of another type. A file whose contents don't match its extension is turned away by the server, and the agent says so.
- Unreadable file → a corrupt document, or one whose text could not be read reliably, is named as such.
- Too many words, or a processing failure → a document over your plan's word limit, or one the service failed to process, gets the answer that the document could not be processed.
- A limit reached → the hourly per-visitor limit, your monthly budget and a plan that doesn't include the feature produce the same answer: that document analysis is not available on this site. This is deliberate: the visitor is not told which of the three applied, so uploading files reveals neither your plan nor your remaining budget. The precise reason is recorded in our audit log and shown in your analytics under Rejections by Reason on the Document Analysis tab.
- Not on a plan that includes the feature, or the monthly budget used up → on top of the above, the paperclip is normally hidden in the first place.
PDF page limit: a PDF of more than 1,000 pages is refused before its pages are read.
Disconnect, and what happens to your data
Switch a template's Active toggle off to keep its settings without letting the assistant use it, or delete it from the Templates list — you confirm first, and a deleted template stays deleted. Your other active templates keep working; when no active template is left, the paperclip disappears from the chat and visitors can no longer upload.
Privacy & retention
- The service keeps no copy of the document. The upload travels inside the chat request, and our edge server may hold a large request in a temporary file while it is being received; the service then parses the file in memory during the chat turn and drops it when the turn ends.
- The structured result and stats — word count, page count, detected language, which parser ran, a confidence score and the analysis outcome (e.g. the quote total or the screening verdict) — land in your analytics. The filename, a screened candidate's name and what a Custom template extracted are kept only for turns your Transcript storage setting keeps (by default, only turns the assistant did not answer confidently). The full document text is not stored.
- Per-upload analytics are kept for the period in the Data Retention Policy, then the extracted detail is erased automatically. Your conversation-level analytics (message counts, topics) are unaffected.
- No training data. Uploaded documents are not added to your knowledge base and not stored as vectors. The AI providers that read them do not train their models on them, under the terms we hold with each provider; the Privacy Policy names those terms provider by provider.
- Audit log records denials (plan-not-entitled, file-too-large, rate-limited, monthly-limit-reached, unsupported type, too many words, and named parser outcomes such as a corrupt file, a timeout or an unreliable read) for SOC 2 compliance. Successful uploads aren't audit-logged — they're operational — and neither is an unexpected parser crash with no identified cause, which is a fault to investigate rather than a denial to account for.
Plan and limits
Document Analysis is included on Pro and Business, not on Starter.
A higher plan raises the limits and adds native reading of complex PDFs and your own template wording.
| Starter | Pro | Business | |
|---|---|---|---|
| Document Analysis | — | ✓ | ✓ |
| Uploads per month | — | 200 | 500 |
| Largest file (MB) | — | 10 | 25 |
| Longest document (words) | — | 20,000 | 50,000 |
| Parsing | — | Standard | Standard, plus native reading of complex PDFs |
| Your own template wording | — | — | ✓ |
| Active templates | — | 3 | Unlimited |
| Uploads per visitor per hour | — | 5 | 20 |
- Uploads / month is one count per website, shared by its visitors, and resets on the same day of the month as your message quota, also on a yearly plan, like your other plan meters (see Billing & plans). You can watch it on the Plan Usage card in Website Analytics.
- Per-visitor rate limit protects that monthly budget from a single visitor burning through it — it's counted per visitor's network address per hour, per website.
- Parser tier decides how PDFs get read. Plans with standard parsing read the text layer locally; plans with native reading automatically escalate complex PDFs (tables, scans, multi-column layouts) to the Gemini native PDF parser. See Supported file types.
- Custom templates (your own free-form prompt) are the “Your own template wording” row of the table; plans without it use the eight built-in kinds listed under The built-in templates.
The table above is generated from the plan catalogue our servers read, so it states what your plan allows rather than what this page was last edited to claim. For how much of this month's allowance is left, read the Plan Usage card in Website Analytics.
Setting up a document workflow and want a hand? Email support@novixagent.ai — happy to help you configure it.