Knowledge Base
Add supported PDF, DOCX, TXT, MD, CSV, JSON, HTML and log-like text files, URLs, web search results or notes.
Understand this layer ↗A knowledge layer, not a black box
Four ways to keep useful context
Knowledge retrieval, semantic indexing, explicit memory and Markdown skills serve different purposes. Use the right layer for the job.
Semantic search
Index supported folders with an embedding model, then search by meaning rather than filename alone.
Understand this layer ↗Persistent memory
Store specific preferences or facts with remember/recall tools, separate from the raw transcript of a previous chat.
Understand this layer ↗Reusable skills
Write Markdown instructions for a topic, include steps and success criteria, and choose global rules for recurring workflows.
Understand this layer ↗Topic teaching
Generate topic-specific rules and Q&A examples from an online model and save them for contextual retrieval.
Understand this layer ↗GitHub skill folders
Install and manage skill repositories with the available tooling; note that repository skills and skills-md instructions use distinct paths.
Understand this layer ↗Choose the right layer
Retrieval is not model retraining
Knowledge and semantic search retrieve stored text; memories store explicit facts; skills supply instructions. Topic teaching builds rules and examples for use in context—it does not fine-tune or change the model weights.
| Layer | What it does | Best for |
|---|---|---|
Knowledge Base | Stores text chunks from documents, notes and URLs. | Research documents and trusted reference material. |
Embeddings | Indexes supported files to retrieve meaning-similar chunks. | Finding a relevant passage without remembering its filename. |
Memory | Stores specific facts and preferences locally in app data. | Stable formatting preferences and recurring context. |
Skills | Supplies task-specific instructions to the model. | Repeatable methods, domain rules and output standards. |
Teaching | Creates topic rules and Q&A examples for contextual use. | Reusable topic context without weight updates. |
A library that can answer backSearch by meaning.
Search by meaning.
Return to the source.
Semantic search uses an embedding model to compare meaning across indexed content. The knowledge retrieval path can then bring related excerpts into the active model context. These operations do not rewrite the model’s neural weights.
- ✓Install the optional embedding model from Features.
- ✓Index the folder you want to search.
- ✓Ask a concept-based question and review returned passages.
- ✓Re-index changed files when necessary; unchanged files may be skipped based on modification time.
SEMANTIC QUERY
"Which notes discuss reducing model latency?"
MATCH 01 · deployment-notes.md
SIMILARITY · conceptually related passage
MATCH 02 · architecture-review.docx
SOURCE · file and excerpt available for review
"Which notes discuss reducing model latency?"
MATCH 01 · deployment-notes.md
SIMILARITY · conceptually related passage
MATCH 02 · architecture-review.docx
SOURCE · file and excerpt available for review
⚠
Know what is stored. App data lives under
%LOCALAPPDATA%\AiDesktop. It includes chat history, memory, knowledge chunks, embeddings, skills and provider settings. Back up deliberately and protect configuration files that may contain API keys. Scanned PDFs may have no extractable text unless OCR is provided separately.