What docAnalyzer doeswith a set of documents.

Getting documents in

A set starts with an upload. Drag files in or browse for them, several at once. Paste the URL of a public page or a file and it is fetched and added. Paste or type text into the clipboard box and it becomes a document with the name you give it. Each file is detected by type and converted to a common internal shape. It is cut into pages or sections along its own structure. Then it is indexed. From then on it is a source a chat or a workflow can be pointed at. The original is kept exactly as sent and is never altered.

Existing libraries come across on paid plans. Connect Zotero or Mendeley once, browse your collections, and import the references you choose into a label. Nothing is rebuilt.

In practice

A closing binder, the folder of signed documents that closes a property or company deal, arrives as forty PDFs and three Word files. Drag the folder in. The scanned attachments are read on upload. Label the set by deal, and it is ready for a Focus chat in minutes. A regulator's guidance published as a web page is added by URL and sits in the same label as the internal policy it will be compared against.

Limits

• Standalone images, audio and video cannot be uploaded. Archives are output formats, not inputs.
• No Google Drive or OneDrive connector today.
• Documents are read-only once in.
• Per-file size, monthly upload and storage caps are per plan, on the plan comparison.

Get your first answer in 5 minutes and how docAnalyzer reads your documents cover the mechanics.

Search across all your documents

Search quality is not the weak link. Document search matches on meaning. Ask how much notice you need to walk away and it surfaces the termination clause, even though you never used the word. When a question names its own target and the answer sits in one place, one search finds it.

The limit is elsewhere. A single search has to decide what to look for before it has read anything. It gets one set of results to work with. "Which of these agreements deviate from our standard terms" means reading the standard first. "Does the risk section contradict management's discussion of results" means holding two parts of a filing side by side. "What changed between the 2023 and 2024 filing" means finding the same clause twice. None of those is a search problem. They are read-then-search-again problems. A retrieve-then-generate tool never issues the second query. The failure is quiet. You get a fluent answer built on whatever the first query happened to reach.

A Focus chat is the mode built for this. Pick the documents, or a label that holds them, and ask. The assistant searches. It reads what came back. It works out what is still missing and searches again, carrying its reasoning from each round into the next. Along the way it can:

• read a section in full after finding it, instead of working from an excerpt
• pull a document's table of contents to navigate it
• run an exact-string search when the question is about a name or a defined term
• count where a phrase appears across the set

It chooses which of those to do next from what it has already read. That is what agentic means in practice.

The set is whatever you give it. One document, forty papers, or a labelled set of a few thousand agreements. It is the same Focus chat with a dataset of a different size. Labels expand to their members when the search runs, so adding a label to a chat adds everything in it. Notes are sources too. A memo you wrote last week can sit in the same set as the filings it is about.

In practice

A label holds sixty vendor agreements and your standard terms as a note. Ask which agreements deviate from the standard terms. The first round reads the note. The second searches each agreement for the matching clause. The third reads in full the clauses the search only excerpted, and the fourth searches again for the four agreements where the clause is titled differently. The answer lists the deviations, each with a reference into the agreement, and the chat shows what the assistant did while it did it. Ask the follow-up: which of the deviating agreements were signed after the standard changed. It searches the same set again rather than starting from nothing.

When to use it, and when not

• Use a Focus chat when the question needs reasoning across documents, or when the answer is not in one place.
• Use a workflow instead when the job is the same task on every document, where a per-document result is the point.
• Use Ask docAnalyzer when you do not yet know which set to open.
• For a question you know sits on one page of one document, a single search finds it. Any tool does that.

Limits

• It is not a search box over your workspace. For "which documents match this description", use Smart Search and Selection, which assembles a set rather than answering a question.
• It does not fix text that was read wrong at upload. A bad scan gives bad search. The OCR section says what to do about that.
• A long conversation eventually sheds its oldest material, by a fixed rule, oldest and least useful first. A question that depends on turn two of a fifty-turn thread is better asked in a new chat with the set attached.
• The monthly credit allowance is per plan and listed on the pricing page. There is no cap on pages in a dataset.

How the search decides it is done, and where a citation points, is on how docAnalyzer grounds its answers. The mechanics of asking well are in phrasing tips.

Citations that open the source at the page

When docAnalyzer states something, it attaches a reference to the place that supports it. The interface turns the reference into a link. Click it and the viewer opens the source at that spot: the page of a PDF, Word file, slide deck or EPUB, the section of a note or a text file, the cell, row or column of a spreadsheet. When the chat is working across several documents, the reference also says which one.

The model is not writing "see page 12" into a sentence and hoping. It emits a reference the application resolves, so the link lands somewhere real rather than somewhere plausible. Models are good at sounding certain about page numbers. This design does not ask them to be. Through the API you receive the reference as plain text, so a citation can be checked by a machine and not only by eye.

Making that land reliably is work done at upload. Each document is indexed along the structure it already has: by page for page-based documents, by heading for notes and text, by location in the sheet for spreadsheets. That map is kept as the single record used for both searching and citing. The thing the assistant searched and the thing your link opens are the same piece. Regulatory filings repeat their headings. A US annual report, the 10-K, can carry several sections all headed "Item 7". Those are told apart before the document is cut, so a citation opens the section it means.

A slider sets how strictly answers stay inside your sources. High keeps the model inside your sources, which is where citation-heavy and regulated work belongs. Low lets it draw on general knowledge, which suits ideation and thin sets. It is a bias, not a filter. High adherence makes an unsupported claim less likely. It does not make one impossible, which is why every claim carries a reference you can click.

In practice

An answer about a loan condition cites a page in the loan agreement. Click it and the viewer opens the agreement on that page with the clause in view. An answer about a figure in an attached spreadsheet cites a cell, and the viewer opens the sheet at that cell. An answer that draws on two annual reports carries two references, each naming its document, so a reviewer knows which one to open. Save that answer as a Note and the references come with it. Open the note in Co-work chat and they survive the editing. Through the API the same references arrive as text, so a script can open each one and confirm the page exists and the quoted text is on it.

When to use it, and when not

• Use them whenever the answer will be checked by someone else: a reviewer, a client, a regulator, a co-author. Citations are the reason to put a name on the output.
• Turn adherence high for that work.
• Turn it low when you are exploring a thin set and want the model's general knowledge in the room. Treat those answers as ideas rather than findings.

Limits

• Precision follows the format, and it follows the text.
• A scanned page whose text was read wrong at upload yields a citation that opens on the right page and quotes something that is not there. That is the sign to run the higher-accuracy OCR pass on that document.

The docs on citations and deep links show what the chat displays and what sits underneath. Tune how strictly answers stick to sources covers the slider.

Reading scanned documents

A scanned PDF has no text layer. On upload, docAnalyzer detects scanned or image-based pages and reads the text off them. It works in more than forty languages. It costs nothing and there is nothing to switch on. The text is indexed and searchable like any other document. A citation into it opens the page.

That automatic pass handles clean typed scans well. Handwriting, low-resolution copies and dense multi-column layouts are harder, and no amount of search quality makes up for text that was read wrong at the start. For those documents, a higher-accuracy pass is available from the document's menu. It re-reads the document and replaces the text. It costs one credit per page and is charged whether or not the result comes out better, which is the honest trade on pages that were difficult to begin with. A citation that opens on nothing, or quotes something that is not on the page, is usually the sign you need it.

Images inside a document are a separate matter. Figures, charts and diagrams embedded in a PDF can be looked at by a model that can see, within a cap per question. That is for a chart in a report, not for processing images in bulk.

In practice

A box of old contracts arrives as two hundred scanned pages. On upload the typed pages come back as searchable text and citations into them open on the right page. Three pages are a faxed table of figures with handwriting in the margins. A question that depends on that table returns a citation that opens on the page and quotes text that is not there. Run the higher-accuracy pass on that document, ask again, and the table reads. The three credits per page were worth it on that document and would have been wasted on the other hundred and ninety-seven.

When to use it, and when not

• Automatic OCR needs no decision.
• Run the higher-accuracy pass on a document only after a citation into it has failed or a search has missed something you know is there. Run it per document rather than across the whole set.
• Do not use either pass as a bulk image pipeline. A folder of photographs is not a document set.

Limits

• Standalone image files cannot be uploaded, and neither can audio or video. A photograph has to sit inside a document to be read.
• Handwriting, low-resolution copies and dense multi-column layouts are the ceiling for the automatic pass.
• The higher-accuracy pass costs one credit per page, charged whether or not the result improves.

The docs on automatic OCR and Enhanced OCR cover both passes.

Extract structured data across a set

The recurring version of document work is pulling the same fields out of every document into one table. Parties, dates and liability caps from two hundred agreements. Sample size, country and funding from forty papers. Base rent and renewal terms from a lease portfolio. Reading to fill a table is the part nobody is paid for.

The Data Extractor takes a label and a schema. The schema can be a list of fields or a plain-language description of what you want. It runs one pass per document, in parallel, and returns a spreadsheet with one row per document and one column per field. Each cell is cited to the page it was read from, so verifying a value is a click rather than a search. Results are saved as each document finishes. Start a run across a few hundred documents and close the tab. When you come back, inspect the results, collect them, or cancel the run.

A run returns its result in one fixed shape, which is rarely the shape you need to send to someone. Open the finished run in a chat and the result loads into a panel with the chat box beneath it. Ask for columns dropped or reordered. Ask for a chart drawn from the rows, the table as a PDF, or several files bundled together. The chat reads the result directly. There is no copy-paste, no re-upload, and no second run against your credits.

In practice

Two hundred leases in a label, and a schema written as a list: tenant, premises, start date, expiry, base rent, rent increases, renewal option, whether the tenant may assign the lease. The run starts, and rows appear as each lease finishes. Close the tab, come back after lunch, and the table is complete. Sort by expiry, and click the base rent in a row that looks low. The lease opens at the rent schedule, where the figure is monthly and the neighbouring rows are annual. Ask the chat below the panel to normalise the rent column to annual and hand it back as a spreadsheet. It does that from the result it already has.

When to use it, and when not

• Use the Data Extractor when the same fields exist in every document and a table is the deliverable.
• Use a Focus chat instead when the question is different for each document, or when the answer needs two documents held together.
• Do not treat the table as final. The citation per cell is there because somebody checks the cells before the table goes anywhere.

Limits

• Every extracted value is a candidate. A person checks it before it enters a model or a filing.
• One fixed result shape per run. Reshape it in chat rather than re-running.
• Workflows are credit-metered.

The extract structured data page walks through defining a schema.

Workflows: the same question, summary or audit across every document

The Data Extractor is one of a family. Each workflow runs the same task across every document in a set and returns one structured result. Each is the right tool for a different shape of job.

Individual asks the same question of every document and returns a per-document answer. Which document supports the position and which contradicts it stays visible, instead of blending into one summary. Summarizer produces one summary per document at the length, tone and format you set, so a corpus gets uniform digests rather than whatever each reading happened to produce. Blueprint audits each document against a reference: a template, a checklist or a list of rules. It reports gaps, matches and mismatches per document. Humanizer rewrites AI-generated prose through targeted edits while holding the facts steady. SEO Metadata generates titles, descriptions and preview tags for a set of pages.

A workflow is non-interactive. You configure it, it runs, and results land as each document finishes. It is the same engine as the chat, in bulk. When a request in Ask docAnalyzer fits a batch operation better than a conversation, "summarize each of my files" or "pull every renewal date from these contracts", it proposes the workflow and you confirm.

In practice

Thirty supplier security reports and your own checklist of requirements. Blueprint runs the checklist against each report and returns thirty compliance reports. Each lists which requirements are evidenced, which are missing and which are contradicted. Each finding carries a reference. Open the run in a chat and ask for one table with the gaps highlighted, one row per supplier and one column per requirement. It builds the spreadsheet from the result it already has. The Individual workflow answers a narrower question the same way: does this report name a subcontractor handling data outside the EU, yes or no, with the passage, across all thirty.

When to use it, and when not

• Use a workflow when the output should be one result per document and the same instruction applies to all of them.
• Use the Summarizer when the digests need to be uniform, which a chat reading each document in turn will not give you.
• Use Blueprint when there is a reference to audit against.
• Do not use a workflow for a question that needs reasoning across the set.
• Do not run one to find out what is in a document you have not read. A Focus chat does that faster and without the credits.

Limits

• One fixed result shape per run, which the chat can reshape afterwards without re-running.
• Workflows are credit-metered, and some need a paid plan.
• For a question that needs reasoning across documents rather than the same task on each, a Focus chat is the right tool.

When to run a workflow draws that line. Ask the same question of every document, summarize a collection and audit against a reference cover each workflow.

Files you can send

An answer does not have to be prose. When the work calls for a file, the assistant builds one and it turns up in the conversation as a download. Spreadsheets, with several sheets if the work needs them. PDFs with real pagination, margins, headers and footers, in the page sizes you would expect. Web pages that carry their own images and styling inside one file, so nothing breaks when you email them. Charts and diagrams drawn from your data. Bundles collecting several of the above.

The assistant decides when a file is the right answer, based on what you asked. There is no menu where you pick "make me a PDF". Phrasing matters. "Put the deviations in a spreadsheet with the clause and the contract name" gets you a spreadsheet. "Tell me about the deviations" gets you prose.

Work accumulates instead of resetting. Every file produced in a chat stays available for the rest of that chat, and the assistant can still reach it. Extract data into a spreadsheet early on and keep working. Eleven questions later, ask for that spreadsheet as a PDF. Or ask for it bundled with a chart you made since, or rebuilt with two columns dropped. The assistant passes the file it already built into the next step. It does not regenerate the content and hope it comes out the same. Any file it has built can be converted afterwards, so a report becomes a Word document and a spreadsheet becomes a PDF.

Structured files are checked before you see them. A spreadsheet, a PDF or a web page is validated against the shape it is supposed to have. One that does not hold up is rebuilt, and if it keeps failing the job moves to a different model rather than handing you something broken. The check runs whether or not the model followed its instructions.

In practice

A contract review chat produces a spreadsheet of deviations at turn three. At turn nine you ask for a memo summarising the material ones, and it arrives as a PDF with page numbers and a header. At turn eleven you ask for both, plus the chart of deviations by clause type made at turn six, as one bundle to send to the client. The ZIP arrives with the original filenames. Nothing was regenerated. The spreadsheet in the bundle is the one you checked at turn three.

When to use it, and when not

• Ask for a file when somebody else is waiting for one, and name the shape you want in the request.
• Do not expect the file to be there next week. Download it, or save the answer as a Note if the text is what you need to keep.
• Do not ask for slides. There is no PowerPoint output. The deck is made in your slide tool from the PDF and the tables.

Limits

• Files live with the chat that made them, up to twelve per chat. They are not filed into the workspace alongside your documents.
• Saving an answer as a Note keeps the text, not the file.
• Spreadsheets carry the usual Excel constraints, including the sheet-name length.
• No PowerPoint output.

How downloads work, reuse outputs across turns and what you can generate cover the details.

Three ways to chat, and Notes

One chat engine runs in three scopes.

Ask docAnalyzer sees the whole workspace: every document, note and label, plus product knowledge. Ask it what is in the workspace, what to look at first, or how credits work. It answers in place. Ask it to summarize the Q3 contracts and pull the renewal dates. It plans the steps and opens the right surface: a Focus chat for analysis on a set, Co-work chat for editing a note, a workflow for a batch job.

Focus chat sees a set and nothing else. It is the mode for digging: the multi-round search, the citations, the files. One document, a label, or any mix of documents and notes is the same Focus chat with a different dataset.

Co-work chat sees one Note. Open a note in the editor and the assistant edits it alongside you, turn by turn: draft a section, tighten a paragraph, turn a list into a table, change the tone. Every turn is small and reversible. The note and the conversation sit side by side.

The three loop. A Focus chat answers a question about the set. Save the answer as a Note, and the citations come with it. Co-work chat refines the note with you. The refined note becomes a source in a new set, or the input to a workflow, or a file you download.

Notes are your writable documents in the workspace. Uploaded documents are read-only. Notes are where drafting happens. The editor has headings, lists, tables and code. A note round-trips to markdown. A note can sit in a dataset next to the documents it is about, so last week's memo is searchable alongside the filings. On Team and Enterprise plans, several people edit a note at once.

In any chat, typing @ inserts a document, note or label by name. Typing / opens the common launches. A searchable list of every shortcut is one key away.

In practice

You open Ask docAnalyzer on a workspace with four labels and ask what would be the fastest way to find every change-of-control clause, the clause that says what happens if a company is sold, across the vendor contracts. It suggests the Individual workflow on the vendor label with that question, and you confirm. The run finishes. You open it in a Focus chat on the same label, ask which of the flagged clauses would be triggered by a minority investment, and save the answer as a Note. In Co-work chat you turn the note into a two-page memo. The citations into the agreements are still there when you download it as a PDF.

When to use which

• Ask docAnalyzer for orientation, planning, and product questions.
• Focus chat for the analysis itself.
• Co-work chat for writing and editing.
• A note for anything you want to keep and reuse as a source.

Limits

• Co-work chat edits Notes, not uploaded files.
• Inserting a name with @ does not add that document to the chat's sources. The dataset does that.

Three ways to chat and when to use which chat go deeper, with the decision table. Capture your own context with notes covers Notes.

Find documents by describing them

A workspace with a few thousand documents needs a way to pick the two hundred that matter before any question is asked. Smart Search and Selection is a button in the documents grid. Type what you want in plain language: "contracts with renewal terms longer than one year", "files that mention employee onboarding", "leases with an exclusive-use restriction". The system scans every document in the workspace and returns the matching set.

Then decide what to do with it:

• add the matches to your current selection
• keep only the matches
• remove the matches
• replace the selection with the matches

Because the four operations compose, you can stack searches. Find one thing, narrow by a second, add a third. Three sentences give you a precise working set. That selection feeds straight into a Focus chat, a workflow, or a chatbot.

In practice

A workspace holds three years of insurance claims files. Type "auto claims involving a rental vehicle" and replace the selection with the matches. Type "claims with a police report attached" and keep only the matches. Type "claims from the northeast region" and remove them. Three sentences and the working set is the one the audit asked for. Open a Focus chat on it, or run the Data Extractor, or save it as a label so it is there tomorrow.

When to use it, and when not

• Use it to assemble a set when you know what the documents are about but not which ones they are.
• Use the grid's filters instead when you know the label, the date range or the filename.
• Use the in-chat search when the question is about what a document says.
• Save the result as a label if you will need the same set again. The selection itself does not persist.

Limits

• It answers "which documents match this description", not "what do the documents say". The second is what the in-chat search does.
• The description is capped at three hundred characters.
• It scans the active workspace only, and the selection resets when you leave the page. Labels are how a set persists.

Pick documents with natural language and find what you've uploaded cover the grid and the filters around it.

Browse and manage the library

Documents, notes, labels and chats each have a grid. Every grid searches by name and filters by chip and by date range. It sorts on several columns. Many rows can be selected at once for a bulk action. The documents grid also filters by label, including the unlabelled, and by page count band. The four-hundred-page filings are one click away from the two-page letters. The labels grid opens a chat on a selection, spawns a chatbot from it, or combines labels. The chats grid filters by what a chat was grounded in: one document, a label, a note, or a mix.

Select a document anywhere and the side panel shows its preview, its metadata and its labels. Select a note and the panel is the editor. Select a chat and the panel shows what it was about and what it was grounded in. Open the document itself and the viewer renders the original, in PDF, web or spreadsheet view. A citation from a chat lands in that viewer at the cited spot.

In practice

Filter the documents grid to the unlabelled and sort by created date. Select the twenty from last week and apply the label for their project. Then filter by that label and by page count over two hundred and fifty. Open the three long ones in a Focus chat.

Limits

• A grid shows the active workspace. Switching workspaces switches its contents.
• A multi-row selection is per session and clears when you leave the page. Labels are how a set persists.

Find what you've uploaded and organize with labels cover the grids.

Work without the mouse

A keyboard layer covers the app. Press G then D, N, L, C or S to go to documents, notes, labels, chats or settings. G then A opens Ask docAnalyzer. One chord opens a new Focus chat. Another jumps to the message box from anywhere. In a grid, single keys rename, label or delete whatever is selected. The arrow keys move the selection. A command palette lists every action by name with its shortcut, so nothing has to be memorised. A cheat sheet lists every active shortcut by section.

Limits

• The bindings are fixed rather than remappable.
• Single-key grid shortcuts are suppressed while you are typing in a field.

Keyboard shortcuts and quick actions lists them all.

Pick the model, keep the grounding

docAnalyzer is not tied to one AI vendor. You choose which model answers a chat, and you can change it in the middle of a conversation, from Anthropic's Claude to OpenAI's GPT to Google's Gemini and others. The chat does not restart and the work already done is not thrown away. If the model you picked is busy or unavailable partway through, another takes over and the answer keeps going rather than failing in front of you.

What matters for trust is what does not move. Retrieval, indexing, the search tools and the citation discipline run on docAnalyzer's side whichever model answers. Changing the model changes the writing. It does not change what the answer is allowed to be grounded in.

On paid plans, a Lab models toggle adds models from smaller and newer providers to the picker, and you can bring your own provider key. With your own key, only the model call runs on your account. Storage, retrieval, workflows and citations stay on the platform. Keys bind per call, so rotating one between turns is transparent. Custom instructions set once in account settings apply to every chat. The chat's advanced settings expose thinking effort, answer length and how much of the set is packed into a request.

In practice

A long analysis is running on one model when the provider starts returning rate limits. The turn continues on the next model in the chain and the answer finishes, with the same citations, because the retrieval did not change. Later you switch the same chat to a different model for the drafting turns, since you prefer its prose. The files built earlier in the chat are still there and still reusable. On a plan with your own key added, the same chat runs its model calls on your provider account. The invoice for the model shows up there rather than here.

When to use which

• The default set for most work.
• The Lab models when you want to test a provider you are evaluating.
• Your own key when your organisation already has a provider agreement whose terms you need to stay under.
• Change the model mid-chat when the writing needs a different hand.
• Do not expect a change of model to fix a bad citation. The grounding did not move.

Limits

• The catalog is curated and changes as models do. The models page lists what is available today with independent quality and speed figures.
• The security page says where each provider operates and what a request contains.
• Premium models draw on the credit allowance with every turn.

Pick a model and set persistent instructions cover the settings.

Workspaces, labels, sharing and chatbots

A workspace is a project: its own documents, notes, labels and chats, counted separately against storage. Switch between them from the picker, and keep one project or one client apart from the next. Deleting a workspace removes its contents.

Labels group documents and notes inside a workspace. They are flat and they overlap, so one agreement can sit in "vendor contracts" and "renewing in Q4" at once, and a label is not a folder. A label is also the unit the rest of the product runs on. Open a Focus chat on it, run a workflow across it, spawn a chatbot from it. Assemble the label before you ask anything.

Everything in a workspace is isolated per tenant and encrypted at rest. It is never used to train models. It is deleted when you delete it. The security page names each processor that touches it.

Share links turn a chat into a URL that visitors open without an account, with an expiry or a password if you want one. They see the conversation, the source documents and the citations. Chatbots go further. Spawn one from a document or a label and set its persona, look and limits. Share it as a snippet embedded in your site, or as a hosted link. Visitors ask questions and get cited answers from the source, no account needed.

In practice

A law firm keeps one workspace per case. Inside the current one, the label "their contracts" holds what the other side sent, and the label "our forms" holds the firm's templates. A Focus chat on both answers the deviation questions. A share link with a password goes to the client for the answer that matters, and the client sees the conversation and can click the same citations. For the product manual the support team maintains, a chatbot spawned from its label sits on the help site and answers customers with cited passages. The label is updated when the manual is.

When to use which

• A workspace for anything that should stay apart: a project, a case, a client.
• A label for a set you will ask about more than once.
• A share link for one conversation one person should see.
• A chatbot for a source many people should be able to question without an account. Never for a source that should not be public.

Limits

• A chatbot's source is fixed when it is created. To point it elsewhere, spawn a new one.
• A chatbot is reachable by anyone with the link, so it suits a handbook or a product manual and not a contract set.
• Per-plan caps on workspaces and labels are on the plan comparison.

Manage workspaces, organize with labels, share a chat link and embed a chatbot cover each.

Teams

On Team and Enterprise plans a workspace is shared. Invite colleagues by email. Each seat has its own login and shares the document library, the labels and the chats. Notes support live collaboration, so two people edit the same draft at once. Billing is per seat, with a minimum seat count per plan, and the credit usage dashboard breaks spend down by workspace and by team versus personal. Enterprise adds single sign-on through your identity provider over SAML, so access follows your own directory.

In practice

Three analysts share a workspace for one project. One uploads and labels. Another runs the Data Extractor and checks the cells. The third drafts the memo in Co-work chat while the first edits the same note beside them. The credit dashboard shows which workspace spent what this month. The team's spend is separated from each person's own workspaces.

Limits

• Team features start on the Team plan. Single sign-on is Enterprise only.
• Seat minimums and per-seat prices are on the pricing page.

Invite teammates, single sign-on and read your credit usage dashboard cover the mechanics.

How credits work

Every plan carries a monthly credit allowance, and credits meter usage: stronger models and their reasoning, larger datasets, workflow runs, the higher-accuracy OCR pass, and the rendering of composed files use more. Everyday chat on the default models over a modest set rounds to zero, and search, citations and automatic OCR never draw on credits. The allowance resets each month. When a month needs more, one-time bundles add credits that do not expire.

A ledger in the account lists every charge, line by line. A dashboard beside it shows where credits went: by day, by month, by reason, by workspace, and split between team and personal on team plans. That is how you find out what one workflow run across a large label cost, or that a colleague has been running everything on the most expensive model.

In practice

A week of contract review on the default model costs nothing beyond the plan. The Data Extractor run across the whole set uses credits, and the dashboard shows the run by reason. The one document that needed the higher-accuracy OCR pass shows as its page count. The bundle bought for that project is still there next quarter.

Limits

• The allowance, the per-action costs and the bundle sizes are per plan and change with the catalog. They live on the pricing page and in how credits work.
• The charts are aggregated. For the individual charges, read the ledger.

Read your credit usage dashboard explains each panel.

API

Everything above is reachable from your own code. Create a key in account settings. Then, over plain HTTP, your code can:

• upload documents and list them
• chat with a document
• run a query across a set
• download the files a query produced

Chatbot endpoints let your own front end talk to a chatbot you configured. Citations arrive as plain-text references, so your code can resolve and check them. There is also an MCP connection for agents you run elsewhere. Its page says what it exposes.

In practice

A nightly job uploads the day's new agreements, runs a query for the renewal terms, and downloads the spreadsheet the query produced into the shared drive. The references in the query response are resolved by the job into page numbers stored next to each value, so the person who reviews the sheet in the morning can open the agreement at the page. When the job hits the plan's upload quota it stops with an error rather than silently skipping files.

Limits

• API keys are available on the Basic plan and above.
• Calls count against the same upload, chat and storage quotas as the app.

Create and manage API keys, the API documentation and connect with MCP have the endpoints.

How to evaluate it against what you use now

One afternoon, on your own documents, no card required.

1. Upload a representative set. Twenty to fifty documents, including a scan or two and one long file. Put them in a label.
2. Ask the question one search cannot answer. Which of these deviate from our standard. What changed between the two versions. Does one part contradict another. Watch whether the assistant searches again after reading.
3. Click three citations. Each should open the page, section or cell. If one opens on the wrong place, note whether the source is a scan.
4. Ask for the file you owe. The deviations as a spreadsheet with the clause and the document name. Then that spreadsheet as a PDF. Open both.
5. Run the same task in bulk. The Data Extractor with your fields across the whole label. Compare the table to what you would have produced by hand.
6. Run steps 2 to 4 in the tool you use today. Same documents, same questions.

Decide from what came back. If the tool you use today answered the deviation question and its citations opened the page, keep it. If it did not, the difference is the product.