Attach a folder of PDFs, notes, and specs to an agent and it stops guessing. Ask a question and it searches your files first, then answers from what it actually found, with the source and page cited. Retrieval runs entirely on your machine: no cloud index, no embeddings API, nothing uploaded.
Your documents never leave your machine. Ingestion, indexing, and retrieval all happen on-device. There is no vector database in the cloud, no embeddings provider in the loop, and no upload of your files, the index is a local SQLite database in your app data. Sensitive contracts, internal wikis, and private research stay private by construction.
A general model knows a lot about the world and nothing about your world, your runbook, your product spec, last quarter's report. A knowledge base closes that gap. LumaBrowser reads your documents, splits them into passages, and indexes them so an agent can pull the exact relevant snippets into its answer, with a citation you can click back to.
Before answering, the agent runs a search over your documents and reasons from the passages it retrieves, so replies reflect your files instead of the model's best guess.
Every retrieved passage carries its source filename and page. Answers come back with clickable citations, so you can verify the claim, not just trust it.
Retrieval uses fast full-text search with BM25 ranking, deterministic, offline, and instant. There's no separate embedding model to download or GPU to spare just to ask a question.
The clearest way to use it today is a per-agent knowledge base: give a specialist agent its own set of documents, and it grounds every answer in them automatically.
In the Agents tab, define a specialist, a policy expert, a codebase guide, a support agent, with its own prompt and tools.
Open the agent and click Add documents. Pick PDFs, Markdown, HTML, or text files from a native file dialog.
Each file is parsed, split into overlapping passages, and full-text indexed on your machine. Re-adding the same file is de-duplicated by content hash.
The agent automatically gains a search tool over its own documents and searches before it answers, returning cited passages.
Because each knowledge base is scoped to a single agent, one agent's documents never bleed into another's. Your HR-policy agent can't see the engineering runbook, and neither can see the other's, unless you add the files to both.
Ask a plain question. The agent calls its search_knowledge_base tool, retrieves the best-matching passages from your files, and answers with the sources attached.
What's our refund window for annual plans?
search_knowledge_baseAnnual plans can be refunded in full within 30 days of the initial charge. After 30 days, annual plans are non-refundable but remain active until the end of the paid term.
Drop in the documents you already have. Text is extracted per page (for PDFs) or per file, then chunked into roughly 500-token passages with overlap so no answer is cut off at a boundary.
PDFs are read page by page so citations can point at the right page. HTML is converted to clean Markdown before indexing.
Everything the UI does is also a plain REST call. Manage an agent's documents over /api/ext/agent-manager, and the agent reaches them through one tool.
| Endpoint | Does | Body / notes |
|---|---|---|
GET /agents/:id/kb |
List the agent's documents. | Returns filenames, page counts, and doc IDs. |
POST /agents/:id/kb |
Ingest documents into the agent's knowledge base. | { "paths": ["/abs/file.pdf", ...] } |
DELETE /agents/:id/kb/:docId |
Remove one document from the base. | - |
search_knowledge_baseThe single tool the agent uses. Args: query (string). It returns the best-matching passages, each tagged with its source file and page, plus the citation source-map the chat UI renders as clickable references. The tool is only offered to an agent once that agent actually has documents, so it never clutters the prompt when there's nothing to search.
See the Knowledge Base API reference for request/response shapes, and AI Sub-Agents for building the specialists that carry these knowledge bases.
Add your documents once and let the agent retrieve what it needs, grounded, cited, and completely on your machine. LumaBrowser is free and runs locally, so your files stay yours. Pair a knowledge base with AI Sub-Agents and Local AI, or head home to see everything it does.