Anyone who has written a non-fiction book or a whitepaper knows the feeling: the real work isn't in phrasing the sentences, it's in the weeks beforehand. Reading studies, conducting interviews, marking up reports, collecting quotes. By the end, all that knowledge sits scattered across a dozen PDFs, three note-taking apps and forty open browser tabs. The only index to this collection is your own head — and your head forgets.
This is exactly where the new Knowledge Repository in Hermes 3000 comes in. It turns your research chaos into a searchable, condensed memory that the AI automatically draws on while you write. And it does so in a way that goes far beyond "yet another folder for files."
The real problem: keeping knowledge available, not just stored
A place to store PDFs doesn't solve your problem. You already have that — it's called a hard drive. The real problem is something else:
- You remember that some study mentioned a number that would fit perfectly into Chapter 4 — but not which one.
- You know two of your sources contradict each other, but you can no longer find the passage.
- You write a paragraph and only notice during editing that your own material would have supported a far more precise claim right there.
Research is only valuable if it's available at the right moment — namely, when you're writing the specific sentence. That's exactly what a knowledge repository delivers when it doesn't just store, but understands and condenses.
Step 1: Condense, don't just store — Karpathy-style
Andrej Karpathy popularized a simple but radical idea: raw data isn't the goal. Value emerges from condensation — distilling large volumes of material down to the few statements that truly matter. We built this principle into the Knowledge Repository in two stages.
Every source is summarized automatically
The moment you upload a source — a PDF, a Word document, pasted text — more happens behind the scenes than you see. The AI:
- summarizes the document and extracts the key points,
- estimates the metadata (author, year, publisher, source type), which you can then correct with a single click,
- splits the text into searchable chunks and computes a semantic representation for each.
An 80-page study becomes a manageable essence — plus the ability to jump back into the original at any time.
"What's in my collection?" — the big-picture view
The second stage goes beyond the individual document. At the press of a button, the Knowledge Repository produces an overview of your entire collection. It answers three questions that would otherwise take days of manual work:
- Core themes — What is your material about as a whole? What threads run through all your sources?
- Contradictions between sources — Where does study A say one thing and report B the opposite? These are often exactly the spots where your most interesting chapter takes shape.
- Gaps & open questions — What's still missing? Where is your argument thin and in need of another source before you publish?
This is condensation at its best: you see not just what you've collected, but what it means. For a whitepaper that has to convince, the gap analysis is worth gold — it shows you the weak spots before a reviewer does.
Step 2: Usable in the background — fuzzy search meets embeddings
Condensed knowledge that you'd still have to search by hand would only be half a solution. The real leap happens where the repository becomes invisible — in the act of writing itself.
Hybrid search: keyword and meaning at once
Behind the Knowledge Repository runs a hybrid retrieval system. It combines two worlds:
- Fuzzy / keyword search finds exact terms, proper nouns and technical jargon — what a classic full-text search can do.
- Embedding-based semantic search finds what you mean, even when different words are used. Write about "employee retention" and it finds the passage about "reducing turnover" — without a single matching word.
Both result sets are fused into a single ranking. So you get precision and depth of meaning, instead of just one or the other.
The AI draws on it automatically while you write
The crucial point: you don't have to copy and paste anything. In both the book chat and text generation, the AI searches your material in the background, pulls in the relevant passages and — importantly — names the sources it relies on. A bibliography emerges from your collection's metadata as a side effect.
One thing we calibrated very carefully on purpose: the AI draws on your material only when it's genuinely relevant. A built-in relevance filter prevents sources from being dredged up for every casual chat question when they don't actually fit. Better no hit than a misleading one — nothing undermines trust in a non-fiction book faster than a misattributed source.
On by default, no setup required
You don't have to manually connect a repository to a book for it to work. By default, the AI draws on all the knowledge you've collected. Only when you want it more targeted — say, because a particular book should rest on one specific collection of material — do you assign individual repositories in the book's AI settings.
Why this especially helps whitepaper authors
Whitepapers live on evidence. A claim without a source is worthless in a B2B context. At the same time, research is often the least-loved phase — dense material, tight deadlines, high accuracy requirements.
With the Knowledge Repository, that balance flips:
- You upload your market reports, studies and internal analyses once.
- The overview immediately shows you the core themes and — crucially — the gaps you need to close before publishing.
- While writing, the AI pulls in the relevant figures and statements and cites them.
- The bibliography is generated along the way.
What used to be a weekend marathon of cross-reading and quote-hunting becomes a guided process in which your own material actively pulls its weight.
Cross-book: your material keeps working
This may be the most underrated benefit. The Knowledge Repository is not tied to a single book. It's your personal, growing research archive.
That means: the material you gathered for your first book is still there for your second. The fifteen studies in your field, the interviews you've conducted over the years, the reports you subscribe to — together they form a foundation that grows more powerful with every project. An expert rarely writes just one book on their subject. With each new publication, your knowledge base deepens, and the AI has access from day one to everything you've ever collected.
Fiction too: the thriller that knows its research
"Knowledge repository" makes you think of non-fiction first. But novels — especially genres like thrillers, historical fiction or science-fiction — stand or fall on the research behind them.
A believable thriller needs accurate knowledge of police work, forensics, weapons, geography, procedural detail. A historical novel lives on accurate details about the clothing, language and daily life of an era. Until now, all that knowledge had to stay in the author's head — or in a separate document that had to sit open beside you the whole time you wrote.
Put that research into a Knowledge Repository and your book knows it. Write a scene in a police station and the AI automatically pulls in your researched material on procedures and hierarchies in the background. The result is scenes that feel right because they're built on real background knowledge — without you having to break your flow to look something up.
How to get started
- In Hermes 3000, open Knowledge Repositories and create a collection of material.
- Upload your sources — PDFs, Word documents, pasted text. Studies, interviews, notes, earlier manuscripts.
- Wait a moment for each source to be processed (the summary, key points, metadata and search index are generated automatically).
- Generate the big-picture view with "What's in my collection?" — core themes, contradictions, gaps.
- Just start writing. In chat and during text generation, the AI draws on your material and cites it.
Research was never the problem — finding it again at the right moment was. That's exactly what your Knowledge Repository now takes care of.
→ Try it now: hermes3000.ai