How Sebtember works

Sebtember is a universal AI assistant platform. You bring your own material — documents, images, spreadsheets, notes, contracts, manuals, price lists, research — from any field, and Sebtember reads it and answers your questions from that material. It's also a way to share trusted content: invite teammates or clients and they get answers from a personal AI assistant that draws only on that shared, trusted knowledge — never the open web. The AI acts like a senior colleague: it reads the relevant material for your question and gives a grounded answer, plus a little more when it genuinely helps.

Three things make it different from a general chatbot:

• It answers from yoursources, not the public internet. If something isn't in the accessible material, it tells you so instead of inventing an answer.
• Answers are grounded in your sources and cited, so you can trace them back.
• You can share your knowledge with teammates or clients, and they can use it for free — you only pay for the answers you generate.

Create your account.Go to sebtember.app, click Sign in (top-right) or Get started, and sign in with your Google account. That's it — there's no setup wizard. You land directly in the chat. Your country and preferred language are detected from your browser; you can change either later on the Profiles page.

Joining via an invitation.If someone sent you an invite link — to a Catalogue (shared content) or a token share (shared balance) — just open the link and sign in with Google. You'll be connected to the shared content or balance immediately.

What you get on day one:a free starting balance of tokens (Sebtember's usage currency) and an empty chat workspace.

The account menu. Your avatar and name sit at the bottom of the left sidebar. Click that row to open Preferences, Profiles, Usage, and Sign out.

Open the account menu → Profiles. There are two tabs. Profilecovers identity (photo, display name, read-only email), localisation (country and preferred language — the AI replies in your preferred language regardless of the source document's language), appearance (light/dark theme), and account deletion. Companyholds company details and VAT (tick "VAT registered", enter your number, and Verify — Sebtember checks it against EU VIES / UK HMRC). Both save automatically as you type.

What the AI knows about you is not here — it lives in Chat settings, described below.

Account menu → Preferences → the CHAT card → the gear. One field: your own brief on yourself. It is the single biggest lever on answer quality, and the AI follows it exactly. Takes effect on your next message.

Write it in whatever shape suits you. ## headings group sections, and an ## Avoid heading is read as real constraints rather than preferences:

• Your role and what you care about."I'm a procurement manager. I care about prices and lead times — keep technical detail minimal unless I ask."
• How you want answers framed."Always cite the exact clause and section. Reply in formal English."
• What to leave out. An ## Avoid section: no marketing language, no motivational framing, no restating the question.
• The tools and standards you work to. Naming them lets the AI use your vocabulary instead of guessing at it.

Sebtember does not build a profile of you.It does not watch how you ask questions, infer your expertise, or keep a picture of you between conversations. What it is told about you is exactly what you write here — nothing more. Earlier versions did infer a profile from your chat history and showed it back to you as "what I've learned about you"; that was removed, along with the behaviour record it was built from.

The one thing it also sees is which catalogues you own — so if you maintain the material on a subject, it can assume you know that subject. That comes from your account, not from anything you have said or done in chat.

Leave the field empty and nothing is lost: answers are still grounded in your catalogues and still cited. The brief changes how they are written, not what they are drawn from.

Click + New chat in the sidebar. Each conversation is its own chat room, listed in the sidebar. The AI remembers the current conversation and keeps a rolling summary of older messages so long chats stay coherent. Its only knowledge sources are the Catalogues you have access to and platform-wide system catalogues (like this help guide) — it does not use the open web, and will tell you when something is outside your sources rather than guess.

The composer. Choose Standard (adjusts depth automatically per question — faster) or Deeper thinking (reasons longer on harder questions — slower, uses more tokens). You can also use Voice input to dictate a message instead of typing, and include an image with your question.

What you'll see in a reply:

• A cited, grounded answer — every factual claim is drawn from your sources and checked against them, so you can trust what you read.
• Sources — the files behind the answer, with page references, a short note on how each contributed, and download links where the file owner allowed downloads. Notes never appear as downloadable sources.
• A "Source:" line naming whose content contributed; click a name to see exactly which of their files were used in that answer.
• A clarifying question — if your request is genuinely ambiguous, the AI asks one focused question with tappable options instead of guessing. Tap one to continue.

Buttons under an answer:

• Copy — copies the answer to your clipboard.
• Share — publishes that single question and answer as a public, read-only page. It is a snapshot, so later private messages in the same chat can never appear on it; sources are cited by name and page without publishing their contents or download links.
• Feedback — sends a private written note to the content owner(s) whose documents grounded the answer. They can reply once, and the reply appears in your chat. They never gain access to your chat room.

A Catalogue is a folder of material the AI is allowed to read. To create one: sidebar → hover CATALOGUES → + New, name it, and open it. Inside a catalogue you'll find a Files tab (documents grouped into Collections), a Users tab (people you've shared with), and an Output tab (catalogue-wide rules for how the AI presents answers).

Uploading files. Click + Add file on a Collection card. There are four types:

• Document (PDF, TXT, CSV, JSON, DOCX, XLS, XLSX, max 25 MB) — extracted and indexed.
• Note (max 10,000 chars) — free text you type yourself; visible to the AI but never shown as a citation or download.
• Image (JPG, PNG, WebP, GIF, max 5 MB) — the AI reads text and content from the image.
• Attachment (max 25 MB) — any file you want downloadable from chat without indexing.

Up to 10 files per upload; drag-and-drop works. Each file shows a live status badge (Pending → Processing → Complete). Per-file settings (the kebab menu) let you add a description, toggle Allow download, or delete. Per-collection settings let you rename, set output guidance, or delete.

Output guidanceshapes the AI's voice. Catalogue-wide rules (Output tab, max 1500 chars) might be "Reply in formal English; show prices including VAT." Per-collection rules (max 800 chars) might be "Contracts: lead with the clause number, then the obligation." Both are optional — without them the AI uses sensible defaults.

Both turn a product name into a written document, drafted from the material in your Catalogues. They are sidebar sections of their own: press + New, type the name and model of one device or material, and press send. There is no chat to have first — the name is the question.

• Specification— a design-stage reference in four parts: bill of materials, specification, installation and operation. Every part is built from your documents' own facts. Where your documents don't cover installation or operation, general best-practice for that type of device is added underneath and clearly flagged as not coming from your files.
• Quotation— a brand-neutral tender specification: numbered chapters, "shall" wording, and quality figures written as minimum or maximum thresholds so suppliers can bid against them. Brand names are left out on purpose.

To narrow a draft, type again in the same document — "only the electrical requirements", or the name of one variant. That replaces the draft rather than adding to it, so an earlier version is not kept. The sidebar row is named after the device, not after your refinements.

If it cannot draft, it says so instead of guessing: when nothing in your documents matches the name, when your documents cover a different product entirely, when you have named several devices at once, when what you typed is not a device or material, or when a refinement is about something else. Each case tells you what to type next, and a refusal leaves the draft you already had.

Both are on by default. You can disable either from Account menu → Preferences, which also carries a short description of each. Disabling hides it and stops it drafting anything new — documents you already have stay readable. Finished documents can be shared as a public page or sent back to the content owner as feedback, the same as a chat answer.

If you used RFQ or BOM as agents before, these are the same two tools with homes of their own — they start from a name you type rather than from a finished answer. Past runs stay where they are in your chats.

Pricing turns a list of materials into a priced one. Open Sidebar → Pricing → + New and upload the list — a PDF, a photo, a spreadsheet or a document. Model numbers are optional: Sebtember works out the specific model for each line from your Catalogues, then prices it from your own price lists.

It tells you how many lines it found before anything is spent, so a mis-read file costs you nothing. You confirm, and the run starts.

The result is a table you can act on, and it is deliberately careful with numbers:

• One column per price list, each keeping the source column’s own heading — a list routinely carries net and gross, so a bare number would be ambiguous exactly where being wrong is most expensive.
• Every total states its coverage(“8/10 priced”) and gives one figure per currency. Two columns are only comparable at equal coverage — the cheaper-looking one is usually just the list that priced fewer lines.
• A line we could not identify says so rather than going blank, and a weak match is flagged for you to check rather than quietly priced.
• Switch price lists off to compare a subset. That is instant and re-runs nothing — every price was already fetched — and the CSV / XLS export follows what you are looking at.

You are billed for the actual run; the estimate shown at the confirm step is only an estimate. A failed run is never billed.

If you used IPL as an agent before, this is the same tool with a home of its own — you no longer have to switch agents on to reach it. Past runs stay where they are in your chats.

A Catalogue invitation shares access to your knowledge so others can ask questions of it. From the sidebar, open the kebab (...) on a catalogue → Invite to this Catalogue, pick a duration (30 min / 1 hour / 24 hours / Unlimited), and send the link. The recipient opens it, signs in, and is added.

Invitees do notsee your catalogue in their sidebar — they only encounter its content as answers in their own chats. Manage everyone on the catalogue's Users tab (pause access or remove). One link can be reused by many people until it expires or you revoke it. Sharing content is separate from sharing tokens (see Tokens & billing): content sharing gives someone the right to ask; token sharing gives them the balance to pay for it.

You can also designate a catalogue as a Library— a standalone knowledge library shown in the LIBRARY section of your sidebar (enable Show LIBRARY in Account menu → Preferences, then use the LIBRARY section's + New).

Beyond invitation links, there are four more ways to reach people:

• Public answer pages. The Share button under any answer publishes that single question and answer as a public page, with contributors credited. All public answers are also listed at /answers.
• A chat widget on your own website. Create a widget from the account menu, attach one or more of your private catalogues, and paste a one-line script tag onto your site. Visitors get a chat launcher, sign in with Google inside it, and ask questions answered from your attached catalogues. Unticking a catalogue or deleting the widget revokes visitor access immediately, and the activity panel shows anonymous statistics only. Widgets can expose private catalogues only — never published Libraries — and visitors pay for their own questions.
• Slack. Install the Sebtember bot into a workspace once and the whole workspace can ask questions — in channels, threads, direct messages, or with the /sebtember command. Each member links their own Sebtember account.
• Other AI assistants. Sebtember can act as a knowledge source for external assistants that support the Model Context Protocol, so they answer from your Sebtember content. Connect one from the account menu, either with a one-click approval or an API key; disconnecting revokes access immediately.

Sebtember runs on tokens. Every AI answer you generate consumes some; you top up to refill. You pay for the answers you generate — the cost is debited when the AI replies. Reading old messages is free; asking is what costs. Content shared with you is free to read from. Platform-wide system content is always free.

You can have two balances: My balance (tokens you bought) and a Shared balance (tokens someone else shared with you). When you have a shared balance, a source selector chip appears in the chat header so you can choose which balance a conversation spends from.

Topping up:account menu → Usage → Billing. Pick a quantity (millions of tokens); the price updates live with VAT for your country. Tick the country-match consent and pay via Stripe's secure checkout — your balance updates within seconds and a VAT invoice PDF is generated automatically. Your card's issuing country must match your profile country; if they differ the charge is voided automatically (no money moves).

Sharing tokens:you can fund someone else's usage up to a monthly cap. On Usage → Billing, click the share icon on your balance card → Create share, choose a monthly cap, and send the link. The recipient accepts and is seeded with the cap from your balance; each month the share tops them back up to the cap, and you pay only for what they actually used. Manage shares under Usage → Analytics → Tokens Users (pause, resume, change the cap, or delete).

Agents, voice transcription, and searches from Slack or a connected assistant are billed the same way — to whoever runs them. Visitors to a website widget pay for their own questions.

The Usage page has three tabs: Billing (balance, top-up, receipts), Analytics (who consumes your shared content, and the people you fund), and Insights— a "content gap" feed showing anonymised questions your users asked that your catalogue almost answered. You see the question and which catalogue it targeted, never who asked, so you can decide what to upload next.

The single most important rule: garbage in, garbage out. The AI reads your documents exactly as you provide them. A blurry scan stays blurry, and a wrong price list produces wrong answers — confidently. Two minutes of prep per file pays back for as long as that file lives in Sebtember.

Maximize accuracy:

• English first. Upload the English version where you have one — it processes faster and avoids double-translation drift on names, codes, and unit conversions.
• No blurry scans. Sebtember reads PDFs visually, page by page. Skewed, watermarked, or photocopied-of-a-photocopy pages produce garbled text and missed columns. If a human has to squint, the AI will misread.
• Tables: Excel and CSV beat PDFs. A native spreadsheet is read as structured data — every cell, every row. The same table inside a PDF is read as a picture.
• One topic per file. A focused 20-page document is indexed better than a 500-page combined-everything PDF, and yields more precise citations.

Save tokens.Uploading a file is a one-time extraction cost; then every answer is paid for at chat-time by whoever asks. Native files (DOCX, XLSX, CSV, JSON, TXT) cost dramatically less than PDFs, which are processed page-by-page through vision AI. Delete junk pages (disclaimers, blank pages, repeated glossaries), don't upload duplicates, and keep files lean — bloated files make every answer slower, costlier, and less accurate for everyone you share with.

What not to upload: image-heavy brochures (mostly cost, little signal), password-protected or encrypted PDFs (extraction will fail — unlock first), drawings or blueprints with no text labels (attach them as downloadable Attachments instead), and fax-quality scans you can barely read yourself.

Images & photos. Use the Image type for screenshots, labels, forms, or photos with visible text. Clear text is mandatory; glare, motion blur, and oblique angles produce garbage extraction. Diagrams are only useful if their parts are labelled in text. Resize anything over 5 MB.

Attachmentsbypass AI extraction entirely (zero processing tokens) and simply create a secure download link — use them for software, large design files, installer packages, and anything you want downloadable but don't need the AI to read. Noteslet you type up to 10,000 characters of context, tribal knowledge, or errata directly into a Collection; they're free to write, visible to the AI, and hidden from your chat users.

File descriptions give the AI a map. A file named doc-2.pdftells the AI almost nothing. The same file described as "2026 master services agreement with ACME — covers termination and payment terms, signed" gets indexed against every term in that sentence. Always include the subject, the document type, and the year or revision. Write in your native language — Sebtember translates descriptions to English behind the scenes.

Organize with Collectionsby content type (e.g. "Contracts", "Price Lists", "Manuals", "Reports"). The AI uses Collection structure to retrieve faster, and you can apply per-Collection output guidance that doesn't bleed into other answers. Finally, the Allow Downloadtoggle: ON lets the AI offer a download link; OFF lets it quote from the file but not hand over the original — use OFF for content you license but don't own.

• A file says Error / Fatal error— usually encrypted, password-protected, or corrupt (unlock and re-upload), over the size limit (image > 5 MB or other > 25 MB — resize/split), or a blocked type (executables/scripts).
• The AI can't find something in your files — the file may still be processing, may be in a Catalogue you no longer access, or scanned poorly (re-scan at 300dpi+). Rephrase with terms from the document.
• You ran out of tokens — top up on Usage → Billing, or ask whoever shares a balance with you to resume or raise your cap.
• Stripe declined / country mismatch— your card's issuing country must match your profile country. Fix one and retry; no charge happens until they match.
• Your VAT number won't verify — type the full number with the country prefix (e.g. LT123456789), click Verify, and refresh after a minute. Both the number and the company name must match the official record.
• Where are your receipts — account menu → Usage → Billing → Receipts. Each completed top-up produces a downloadable VAT PDF within a minute.

The account menu is your avatar and name at the bottom of the left sidebar.

Your data stays yours. Files are stored privately, encrypted, EU-hosted on Google's enterprise infrastructure, and never used to train AI models. Sebtember never uses the open web to answer, and deleted data is permanently wiped. For the full picture, see the Security page, the Privacy Policy, and the Terms. Questions: hello@sebtember.app.