Back to blog
Product5 min read

Lose Nothing: The Parts of Your Documents Most AI Tools Drop

By Vaidas Saltenis · Founder, Sebtember

Ask an AI tool to read a document and it will almost always give you an answer. What it rarely tells you is how much of the document it actually understood.

Most tools read the easy part — the flowing paragraphs — and quietly drop the hard parts: the pricing table, the dimensioned drawing, the performance curve, the clause written in another language. You don't get an error. You get a confident answer built on a fraction of the file.

Sebtember is built on the opposite principle. Lose nothing.

What gets lost — and what Sebtember keeps

Most tools read the flowing text and quietly drop the parts that are hardest to read. Here is how the tools people compare — NotebookLM, Claude, OpenAI, and the specialist parsers — handle each dimension of a document, and what Sebtember does instead. Scroll the table sideways to see every provider.

strong partial weak or absent not applicable to that product’s shape
DimensionSebtemberthis pipelineNotebookLMconsumer, managedClaudeprojects + filesOpenAIfile_searchSpecialistsLlamaParse · Reducto · Azure DI
What it can read
Scanned-PDF OCRCloud Vision per page; per-page confidence surfaced as a warningManaged, opaqueAnswer-time page vision; no stored OCR layerWeak on scansBest-in-class
Digital tables → retrievable structurePer-row vectors + structured JSON + row-level citationStructure-lossy, no row indexRead from the page at answer timeFlattened into prose chunksSchema + bbox tables
Multi-column reading orderGutter detector → VLM; measured on digital and scanned fixturesOpaqueVisual reading handles itText-order dependentDedicated reading-order models
Technical drawings — value ↔ meaningPaired at ingestion and stored (“overall height: 737.1 mm”)Summary onlyAt answer time, per queryNumbers without referentsTranscription, not interpretation
Charts & curves — data-point readingAxes + units + per-curve points, at ingestionDescriptiveAt answer timeVaries by vendor
How it gets indexed
When comprehension happensIngestion, once — and the interpretation is itself retrievableManaged; long-context at answer timeAnswer time, every queryIngestion, but shallowIngestion — output is a file, not an index
Chunking sophisticationContextual enrichment + boundary bridges + heading-aware splitsLong-context stuffing insteadManagedFixed 800 / 400 defaultChunking is yours to build
Visual-vector retrieval (page embeddings)Multimodal page vectors in the same index as text, in productionEmerging
Hybrid lexical + dense retrievalVector + keyword + rank fusion + cross-encoder rerankNo exposed retrieval layerManagedDense-ledNot a retrieval product
MultilingualOriginal-language storage + bilingual keywords → verbatim exact-matchStrongStrongDense-only — no verbatim matchStrong
Corpus-level synthesisRollup vectors, so “what’s in here?” has a retrieval targetLong context sees the whole sourceContext-window boundChunk-level only
Whether you can trust the result
Failure honestyWarnings taxonomy on the document — which pages read poorly, which caps were hitA source just fails genericallyRequest-level errorsFile status APIJob status APIs
Self-healing after a hard killBounded auto-resume from checkpoint; the recovery path is itself testedManaged, invisibleManagedManagedManaged
Ingestion observabilityLive page-by-page progress + ETA, per-file token breakdown, routing per pageA spinnerStatus pollingJob status
Ingestion-side eval harnessGolden fixtures run against the deployed function; claims are measured, not assertedImpossible by constructionImpossible by constructionImpossible by constructionNot shipped
Prompt-injection hardening at ingestionDocument text is treated as data; structural tags stripped before indexingNot disclosedModel-levelModel-levelParsers don’t prompt
What it costs and where it runs
Per-file cost attributionItemized, per-payer, debt-aware — what this document cost, and who owes itSeat pricingSeat pricingStorage + query meteringPer-page pricing
Scale ceiling per document25 MB upload · 540 s per run · checkpoint cliff on the largest PDFsVery large per sourcePer-request page limitVery large per fileUnbounded per-page jobs
Source sync & connectorsManual upload onlyDrive, web, YouTube, audioConnectors, tier-dependentBuild your own
Data residency & tenancyRuns in a project you control, in a region you chooseVendor-managedEnterprise optionsEnterprise optionsSome offer on-prem

The Sebtember column is verified against the deployed pipeline. The other columns are our reading of each product in mid-2026; these tools evolve quickly, so treat them as directional.

The row that matters most is "when it reads the file." Reading the page on every question sounds fine until you realise you cannot search by something that was never written down. Sebtember reads once, at upload, and stores the interpretation — so the depth is already there the moment you ask.

We have yet to find another tool that keeps all of this. Most keep the prose and quietly let the rest go.

How it works: many technologies, one job

No single model reads every kind of page best — so Sebtember doesn't rely on one.

Each page is routed to the tool that fits it. A clean digital page uses its own native text layer. A scanned page goes through dedicated OCR. A diagram or a complex table goes to a vision model that interprets it rather than merely transcribing. The results are embedded with best-in-class models, indexed for both meaning-based and exact-keyword search, and re-ranked so the right passage rises to the top.

It is a pipeline of many specialised technologies and providers, each doing the one thing it does best — and it is constantly evolving. Extraction quality isn't a fixed feature; it's a moving target we keep chasing. When a better engine, model, or technique appears, it goes into the pipeline. So the promise isn't "the best extraction of last year." It's the best result available on the day you upload.

How do you know it actually worked?

A promise to lose nothing is worth only as much as your ability to check it. So Sebtember does two things most tools don't.

First, it tells you how well it read your file. Instead of a spinner that ends in "done," you get live page-by-page progress and, when the file finishes, a plain record of anything that went less than perfectly — which pages scanned poorly, where a limit was reached. "Your file is ready, and pages 12–14 were low quality" is a far more honest answer than "your file is ready."

Second — and as a company built on trustworthy answers, this is the part we care about most — the claims in this article are measured, not asserted. The extraction pipeline is checked by golden-set tests that run against the actually-deployed system: known documents in, known values expected out. "It reads tables correctly" isn't a slogan here; it is an assertion that passes or fails on every change we make.

And every answer traces back to the exact page, the exact row, and the exact passage the model saw — so you are never asked to take the result on faith.

The bottom line

Most AI tools lose the hardest, most valuable parts of your documents and never mention it. Sebtember is built to lose nothing — to read the table as a table, the drawing as real measurements, the chart as real data, your language as your language — to store all of it so it stays findable, and to show you how well it did.

Because an answer is only as trustworthy as the reading behind it. You should never have to wonder how much of your own document the AI actually saw.

Upload the file everything else struggles with — the dense price list, the dimensioned drawing, the contract in two languages — and ask it the question you actually care about.

Frequently asked questions

Why do AI tools lose tables and data when reading PDFs?
Most tools flatten a table into flowing prose and only read the page at the moment you ask, so the column relationships and exact values are lost — the table stops being something you can query row by row. Sebtember instead turns each row into its own retrievable record with its own citation, so a question like 'what does SKU 1800 cost?' returns the exact line, not just the right document.
Can AI extract exact values from a technical drawing?
Most tools transcribe the loose numbers on a drawing without recording what each one measures, so the values can't be retrieved by meaning. Sebtember pairs each measurement with its label as the file is ingested — 'overall height: 737.1 mm' — and stores that pairing, so you can ask for the height and actually get it.
Does Sebtember read scanned PDFs, images, and charts?
Yes. Scanned pages go through OCR, images and diagrams go through a vision model that interprets them, and charts have their axes, units, and data points read off — all stored in the same searchable index as the text.
Does it work with documents that aren't in English?
Yes. Documents are stored in the language they were written in, with an English keyword layer added alongside rather than replacing them. A verbatim quote pasted from your own non-English document matches exactly, instead of being approximated by a translate-first tool.
How do I know the extraction was accurate?
Sebtember surfaces a per-file record of anything that read poorly (for example, low-quality scanned pages), and its extraction is checked by golden-set tests that run against the deployed system — measured, not asserted. Every answer also traces back to the exact page, row, and passage the model saw.