Product1 Jul 2026
Upload once, use everywhereThe wiki
The wiki
your AI reads.
Sebtember turns your documents, price lists, drawings and notes into clean, AI-friendly knowledge. Upload a file once: ChatGPT, Claude, Plugent and your own tools all read the same source, with citations.
1,500,000 free tokens to start · no subscription · hosted in the EUPDFXLSXJPGDOCX
Heat pump manuals62 files · ready for AIReadyReadyReady
ChatGPTClaudePlugentYour website
The problem
Messy data burns AI resources.
AI tools are only as good as what you feed them. Today you feed them raw files, again and again, in every tool.
The same upload, again and againEvery new chat and every new tool starts from zero. You attach the same manual, price list and contract over and over.
Raw files waste tokensScans, drawings and big spreadsheets are re-read page by page every time you attach them. You pay for that work again and again, and answers still miss the table on page 84.
Five tools, five copiesEach AI tool keeps its own copy of your files. None of them knows which version is current, and nobody knows who can see what.
The solution
How it worksEvery person and every organisation gets its own AI-friendly wiki.
Sebtember reads each file once: text, scans, tables, drawings and price lists. It keeps the result as clean, searchable knowledge in your catalogues, so any AI can find the right passage and cite it.
1Upload onceAdd documents, images, notes and spreadsheets to a catalogue: even scans, drawings and price lists.
2Extracted onceSebtember reads every page, keeps tables as tables and turns it all into searchable knowledge. You see each file's status, and pages that were hard to read are flagged.
3Used everywhereEvery connected AI tool searches the same catalogue and cites the file it used.
What you get
One source. Every agent, task and prompt.
No more repeated uploadsExtract a file once. Every chat, agent and task after that reads the result instead of the raw file.
One source for every AIChatGPT, Claude, Plugent agents and your own prompts all answer from the same up-to-date catalogue.
Answers you can checkEvery answer points to the file it came from, so anyone can open the source and verify it.
You decide who reads whatChoose which catalogues each person and each tool may use. Pause or disconnect any of them in one click.
Library
Trusted sources you never have to build.
Standards, manufacturer manuals and company handbooks, prepared by the people who own them. Join one with a link and your AI tools use it straight away, next to your own catalogues.
Nothing to upload. The owner prepares and extracts the files, so you don't spend a token or an hour on it.Always current. When the owner updates a standard or a manual, every user gets the new version.Trusted by design. Each Library catalogue shows who published it. Answers cite its files like your own.Works in your tools. Library catalogues reach Plugent, ChatGPT and Claude next to your own catalogues.
By Erik Lindqvist · 210 filesLibrary
By Fjällvärme AB · 38 filesLibrary
By Erik Lindqvist · 20 filesLibrary
Works with
Connect the tools you already use.
Sign in with Sebtember from ChatGPT or Claude, add it to any MCP client, put a widget on your website, or use Plugent's ready-made agents. You choose which catalogues each tool may read.
ChatGPTSign in with Sebtember and ask from any chat.
ClaudeAdd Sebtember as a connector in a minute.
PlugentReady-made agents for chat, pricing and troubleshooting.
MCP clientsCursor, VS Code and any tool that speaks MCP.
Website widgetLet visitors ask your catalogues on your own site.
SlackDM it, @mention it or type /sebtember.
Extraction
How Sebtember reads your files.
Good answers start with a good read. Every page goes to the reader that suits it, and every gap is shown, not hidden.
Text pageScanDrawingThe PDF's own textWhen the page has a text layer we can trustpdf.js · trusted-text check · tabular-geometry signatureGoogle Cloud VisionOCR safety net for scanned pagesper-page confidence · flagged below 0.70Gemini vision modelScans, photos, drawings and tablesgemini-3.6-flash · ≥ 300 path ops → VLM · 32,768-token outputVoyage AITurns every passage, and every image, into searchable meaningvoyage-4-large · 1024-d
voyage-multimodal-3.5 · 1024-d
rerank-2.5-liteSearchable knowledgeProse chunks · 2,500 chars, 150 overlaptable_row + structured_data vectorsCosine vector index + keyword indexextract.warnings per file
voyage-multimodal-3.5 · 1024-d
rerank-2.5-liteSearchable knowledgeProse chunks · 2,500 chars, 150 overlaptable_row + structured_data vectorsCosine vector index + keyword indexextract.warnings per file
3 readersRouted page by pageEach page goes to the reader it needs, so a scan, a drawing and a typed page in one PDF are each read the right way.8-page batches3 parallel VLM batchesgzip checkpoints in GCS8 resume attempts
0 → 24 rowsTables stay tablesRows and headers are kept, and a sheet holding several tables is split into each one. In one measured file, a drawn table that plain text reading missed became 24 searchable rows.≤ 2,000 row vectors / file≤ 12 table segments / sheetunique-header normalisation
2,889 pagesMeasured on real documentsPage routing was measured on 199 real documents: 8.8% of pages need the vision model for drawn structure. Every change to extraction must pass the test suites first.eval:extractioneval:drawingeval:vision-modelsvalue recall + table pipeline
1024-dSearch by meaning, images tooEvery passage and every image is embedded with Voyage AI. A search runs meaning and keyword retrieval together, fuses them, and re-ranks the result.voyage-4-largevoyage-multimodal-3.5cosine kNN, top 30Reciprocal Rank Fusionrerank-2.5-lite
Any languageStored as writtenContent is stored, embedded and keyword-indexed in its original language, with an English keyword layer on top, so a question in English still finds a German manual.native-language storageEnglish keyword layerquery rewrite
No hidden gapsHonest about what it couldn't readPages that were hard to read, or content that was cut, are flagged on the file, and the AI is told too, so it can say so.lowOcrConfidencePagestruncatedBatchesblockedBatchescontentTruncated<source_quality>
From the blog
All articlesField notes on AI‑ready knowledge.
Pricing
No seats. No subscription.
Pay only for the tokens you use: once to extract a file, then a little for each search a tool makes. One wallet works for Sebtember, Plugent and every connected tool.