Knowledge base
Your second brain — searchable, with sources.
Import your knowledge and CompanyWizard answers questions from it — with a reference to the source. That is RAG: the model answers from your documents, not from its memory.
Why an LLM on its own is not enough
A language model does not know your company. Ask it about internal rules and it guesses a plausible answer — sometimes right, sometimes invented. These "hallucinations" are the most expensive property of generic AI.
What is missing is the link to your own content — and evidence for where a statement comes from.
Retrieval-augmented generation
On import, every document is split into sections and stored as a vector — via sqlite-vec, a vector search directly inside SQLite, with no external database. For each question the matching sections are retrieved and handed to the model as context.
The answer is built from exactly those sections and cites their source. For large corpora you can add Qdrant; BM25 serves as the leanest fallback.
How an answer comes about
-
Import
Documents are split into sections and indexed as vectors.
-
Retrieve
The semantically closest sections are found — sqlite-vec locally, Qdrant for large corpora, BM25 as fallback.
-
Answer
The model phrases the answer from exactly those sections.
-
Cite
Every statement points to its source — verifiable rather than guessed.
File dump or curated knowledge base?
| Criterion | File dump | Curated knowledge base |
|---|---|---|
| Time to value | Very fast — upload and done | Slower — import, review, structure |
| Answer quality (RAG) | Variable, depends on parsing | High, from clean, precise sections |
| Control over the AI’s view | Low, content vanishes into vectors | Full, Markdown is plain-text checkable |
| Maintenance & correction | Awkward (original, export, re-upload) | In the note, re-indexable at once |
| Versioning / diff | Practically impossible | Plain text: who/when/what is traceable |
| Auditability | Unclear what is indexed | A finite, checkable corpus |
| Scaling to mass data | Higher | Limited, built for high-value knowledge |
| Technology independence | Tied to index/format | High, index re-buildable any time |
What this means for you
- Answers with sources — hallucinations become verifiable.
- Local vector search without an external database (sqlite-vec); Qdrant optional.
- Six import sources: Word, PDF, HTML, Confluence, SharePoint, Mail.
- PII can optionally be pseudonymised before indexing as well.
Honestly: the limits
-
Curation takes effort.
That is the biggest drawback. The file dump clearly wins on time to value — upload and start, instead of tidying up first. Anyone who wants a result in five minutes feels every curation step as friction.
-
Conversion can lose content too.
The step from PDF or Word to Markdown can damage tables, diagrams or layout-dependent meaning. The point of Markdown is not that nothing is lost, but that losses become visible and fixable.
-
Where layout carries meaning, the file is more faithful.
For engineering drawings, balance sheets, forms or presentations, multimodal models read the original visually more accurately than plain text. For such content, keep the original alongside.
-
A copy can go stale.
If the truth lives elsewhere (Confluence, SharePoint) and you copy to Markdown, a fork appears that can drift. Source references, review dates and maintenance agents push back, but it stays a task.
-
Manual curation does not scale infinitely.
The approach shines for high-value knowledge in a manageable volume. It is not the tool to index your entire 2-terabyte file server — and it does not replace an archive system for legally relevant originals.
Sounds right for your organisation?
In a 30-minute demo we show the concept on your concrete use case — no obligation, with concrete suggestions at the end.