Obsidian's interface to lilbee, a local AI search engine: find, run, and manage local AI models, and search your vault and the sites you crawl with them.
Browse a model catalog and pull a model that runs on your own machine, then ask questions about anything in your vault (notes, PDFs, ebooks, code, scans, 150+ file types in all) and get answers with the source one click away.
Everything runs on your computer. Your notes stay in your vault; nothing leaves unless you ask it to. Cloud models are optional, and only when you choose one.
Heads up: it downloads to your computer. The lilbee server (a few hundred MB) on first launch, plus whatever models you pick from the catalog (a few hundred MB up to several GB each). It all stays local.
Point lilbee at a running Ollama server and pick one model for embedding and another for chat from the Hosted tab. Add a PDF, watch it index, then ask a question and get a cited answer, served end to end by Ollama.
With your own key, hosted frontier models show up under Hosted too. Pick a free-tier Gemini model for chat, keep embedding local, and the answer comes from Gemini while still citing your own documents.
The same as Ollama, with LM Studio's local server: pick its embedder and chat model from the Hosted tab, index a PDF, and get a cited answer powered end to end by LM Studio.
Ask a multi-part question in one shot, a bulb part number and the engine's firing order, from different sections of a manual. A local model answers both and cites each.
A scanned, image-only PDF read by a local vision model. The Task Center streams the OCR page by page, then a cited answer reads the support number and publisher straight off the scanned cover, a detail only OCR could surface.
The plugin's own source code added to the library on an Apple M1 Pro: the Task Center embeds all 58 files beside the live GPU placement view, then what is lilbee for Obsidian? gets a cited answer and the citation opens the README at the source.
Add a PDF from the command palette, watch the Task Center index it, then ask a question and get a cited answer. Click the citation and the source preview opens at the exact page.
Crawl a Wikipedia page into your vault and ask it a question, then jump from the citation straight to the cited section of the rendered source.
Crawl a whole site one link deep into an empty vault, hundreds of pages streaming into the explorer as they index, then ask one multi-part question that only makes sense across all of it. A local model answers with reranking and cites several of the crawled pages.
Browse the model catalog without leaving Obsidian: Chat, Embed, Vision, and Rerank tabs, each pulled live from Hugging Face Hub. Models the bundled engine can't run are flagged before you pull.
Search the catalog for a small chat model, watch the download stream start to finish in the Task Center, then activate it, switch to Chat mode, and use it, the full pull-and-use loop without leaving Obsidian.
Four roles on one rail: chat writes the answers, embedding indexes your notes, vision reads scanned PDFs, reranking sharpens the results. Hover each to see what it does, then flip between Search and Chat mode.
On a box with three A100s, the placement view shows each card and what's running on it. A 235-billion-parameter chat model splits across all three, with the embedder mirrored onto every card. Ingest the codebase, watch each card's load move live, then ask a grounded question and get a cited answer streamed from the 235B, opening the exact source it came from.
Brand-new on a fresh vault: install lilbee from the community plugin store, walk the setup wizard (pick a chat model and an embedder), then ask your first question and get a cited answer from your own notes, with a click-through to the source.
The palette as an async control surface: fire a crawl, a file add, and a model download back to back without waiting, watch all three run at once in the Task Center, then ask the just-crawled page a cited question.
50+ settings: search depth, reranking, sampling, parsers, the wiki. Sane defaults; tune the moment you want to. Long model names and large sizes stay on aligned rows.
lilbee is in active development, with frequent releases; updates arrive through Obsidian's plugin updater. view the store listing →
Coming from Google's NotebookLM? lilbee does the same job on your own hardware. see how it compares →
It's early days. If it's useful to you, a ★ on GitHub helps other Obsidian users find it, and bug reports are very welcome.
- a chatbot in a browser tab doesn't know what's in your vault
- Obsidian's search finds the word, not the answer
- when something matters, you open the document and read it yourself
- your notes and your chatbot live in separate apps
- ask in plain English; get a real answer
- every answer points to the exact line it came from; one click to see it
- it works offline, on your computer
- it's right there in Obsidian, next to your notes
ask your vault anything
Type a question in plain English; lilbee reads your notes and files and answers it. It's like having your own private Encarta, built from everything you've collected.
every answer shows its work
Each reply comes with footnotes. Click one and the exact spot opens, right where it came from, so you can trust the answer or check it for yourself.
it reads more than notes
PDFs, ebooks, spreadsheets, code, even scanned pages and photos. Over 150 file types in all. If it's in your vault, lilbee can search it.
pick a model, no account needed
Browse a built-in model catalog, straight from Hugging Face Hub: featured picks up top, or search the full list. Download one with a click; it runs on your computer. 50+ settings to tinker with, good defaults if you'd rather not.
private by default
Your files, the models, the search: it all stays on your machine and works offline. Want a cloud model for one job? Plug one in, and the plugin tells you whenever it's in use.
remembers what you tell it
Turn on memory and lilbee holds onto durable facts about you and how you like your answers, then recalls the relevant ones in later chats, no matter which conversation they came from. Off by default, managed from a Memories view, and never mixed into your citations.
writes a wiki of your knowledge experimental
lilbee can draft linked wiki pages from your library; they land in your vault as ordinary notes and show up in the graph view, alongside your own.
one box, many GPUs, no flags
Got more than one graphics card? lilbee finds them all. The placement view draws every GPU and what's running on it, chat, embedding, vision, reranking, and lets you spread a model across cards or pin a worker to one with a click.
Spread one model's layers across several cards, or pin embedding, vision and reranking workers to the GPUs you choose, then apply and the fleet rebuilds to match. The usage bars update live, so you can see each card's load in real time. On a single machine it stays simple: one card, everything together, nothing to configure.
Do I need Ollama or LM Studio to use lilbee in Obsidian?
No. lilbee downloads and runs the AI models for you; it is a complete model manager, so there is no separate runner to set up. If you already use Ollama or LM Studio, you can point lilbee at them instead.
Does my vault leave my computer?
No. Indexing and search run on your own machine, and your notes stay on disk. lilbee uses a cloud model only if you pick one.
What can it search in my vault?
Your notes and markdown, plus PDFs, code, ebooks, and scanned images through OCR, and whole websites you crawl into the vault. Over 150 file types, with answers that cite the exact source.
Does it work offline?
Yes. With local models in place, lilbee searches, asks, and chats with your vault with no internet connection.
How do I install it?
From the Obsidian community plugin store: open Settings, then Community plugins, search for lilbee, and install. lilbee sets up the models for you on first run.