lilbee — terminal-first local RAG for files, code, and the web

Project site · PyPI · Obsidian plugin

A terminal-first local RAG and search engine for your own files, code, websites, and scanned documents. One install, no sidecar services, fully offline by default.

⚠️ Beta software

lilbee is in active beta development. Every release on PyPI is a pre-release; you must use --pre (or uv's --prerelease=allow) when installing. Interfaces, command names, and on-disk formats may shift between betas. Feedback, bug reports, and issues are very welcome — that's the whole point of the beta.

Latest pre-release (always): lilbee on PyPI →

• Why lilbee
• Previews
• What you can do with it
• TUI
• Hardware requirements
• Install
• Agent integration
• HTTP Server · API reference
• Interactive chat
• Supported formats
• Experimental

Why lilbee

Local AI tools have gotten great at getting you to a chat window fast. The first evening with a local model is genuinely fun. What makes it more than a novelty is grounding: the model needs context from your notes, your files, your codebase. Without that, the local AI tool runs out of places to go.

Local AI can be made more substantial than a chatbot. lilbee lets you pair the chatbot with a real search engine reviewing a curated set of documents. Make a library of what matters to you, let a local model reason over it, and get answers with citations you can click back to the source. Now the model knows your world.

To acheive this in the past a user would manage a background daemon, a separate inference server, model files fetched by hand from the web, and a retrieval layer glued on top. lilbee bundles all of it into one install. Everything lives in one process, in the terminal; including a GGUF model catalog browser.

The same executable ships a Textual TUI, a REST API, an MCP server for AI agents, and a Python library. It runs globally by default, or per-project by dropping a .lilbee/ next to .git/, the same pattern git uses. Curated documents with topic-specificity produce better answers than a single catch-all vault of personal documents, white papers, instruction manuals, codebases, and so on.

An Encarta 99 you build for yourself, from your own files, shaped to your needs.

Previews

Real terminal recordings coming soon. Previews below give the shape of each screen. Written walkthroughs are under docs/benchmarks/: Godot level generator and vision OCR model comparison.

Chat. The default screen. Streaming replies with clickable citations.

 ┌─ lilbee ──────────────────────────────────────────────────────┐
 │ [Chat]  Catalog  Status  Settings  Tasks         [INSERT]     │
 │                                                               │
 │ You:    what does the oil pressure warning mean?              │
 │                                                               │
 │ lilbee: The oil pressure warning indicates low oil            │
 │         pressure.[1] When the light stays on, stop the        │
 │         engine immediately.[2]                                │
 │         ─────────────────────                                 │
 │         Sources                                               │
 │         [1 owners-manual.pdf:42]   <- click to open           │
 │         [2 owners-manual.pdf:43]                              │
 │                                                               │
 │ ┌───────────────────────────────────────────────────────────┐ │
 │ │ Ask anything...                                           │ │
 │ │                                                           │ │
 │ │ Chat: [Qwen3 0.6B v]    Embed: [Nomic v1.5 v]             │ │
 │ └───────────────────────────────────────────────────────────┘ │
 │ SYNC vault   [============------------]  42%                  │
 └───────────────────────────────────────────────────────────────┘

Task Center. Every background job (sync, crawl, wiki build, model pull) in one place. Global concurrency cap; new tasks queue when full.

 ┌─ Task Center ─────────────── [cap 3/3] [Clear]┐
 │ ACTIVE (2)                                    │
 │   [============-----]  42%  PULL  Qwen3 8B    │
 │   [======-----------]  18%  SYNC  vault       │
 │ QUEUED (1)                                    │
 │   CRAWL  https://docs.example.com             │
 │ COMPLETED                                     │
 │   [v] SYNC  vault                  2 min ago  │
 │   [x] PULL  mistral                5 min ago  │
 │   [v] ADD   cv-manual.pdf         12 min ago  │
 └───────────────────────────────────────────────┘

Wiki. Auto-generated concept and entity pages, with drafts awaiting review.

 ┌─ Wiki ────────────────────────────────────────┐
 │ Filter pages...                               │
 │                                               │
 │ Concepts (8)                                  │
 │   Braking Systems               5 src         │
 │   Cooling System                2 src         │
 │ Entities (12)                                 │
 │   Henry Ford                    3 src         │
 │ Drafts (2)                                    │
 │   Tire Pressure                 1 src         │
 ├───────────────────────────────────────────────┤
 │ ┌─ Braking Systems ────────────────────────┐  │
 │ │ 5 sources | faithfulness 0.92            │  │
 │ │                                          │  │
 │ │ Modern braking systems combine hydraulic │  │
 │ │ actuation with ABS to prevent wheel      │  │
 │ │ lockup under heavy deceleration.[1]      │  │
 │ │                                          │  │
 │ │ [1 brake-primer.pdf:8]  <- click         │  │
 │ └──────────────────────────────────────────┘  │
 └───────────────────────────────────────────────┘

Model catalog. Browse, install, and switch roles without leaving the terminal. * indicates the developer's recommendation for each role.

 ┌─ Model Catalog ───────────────────────────────────┐
 │ [All tasks v] [All sizes v] [Featured v]          │
 │ search...                       [Grid | List]     │
 │                                                   │
 │ Our picks                                         │
 │ ┌─────────────┐ ┌─────────────┐ ┌─────────────┐   │
 │ │ Qwen3 0.6B *│ │ Nomic v1.5  │ │ BGE Rerank  │   │
 │ │ [chat]      │ │ [embed]     │ │ [rerank]    │   │
 │ │ [GGUF]      │ │ [GGUF]      │ │ [GGUF]      │   │
 │ │ 450 MB  ok  │ │ 274 MB  ok  │ │ 1.2 GB      │   │
 │ │ [Use]       │ │ [Use]       │ │ [Pull]      │   │
 │ └─────────────┘ └─────────────┘ └─────────────┘   │
 │                                                   │
 │ Chat                                              │
 │ ┌─────────────┐ ┌─────────────┐                   │
 │ │ Qwen3 8B    │ │ Phi-4 14B   │                   │
 │ │ [GGUF]      │ │ [GGUF]      │                   │
 │ │ 4.9 GB      │ │ 9.1 GB      │                   │
 │ │ [Pull]      │ │ [Pull]      │                   │
 │ └─────────────┘ └─────────────┘                   │
 │                  [Load more]                      │
 └───────────────────────────────────────────────────┘

What you can do with it

A personal encyclopedia of what you've collected

Point lilbee at a folder of PDFs, notes, ebooks, or code and it indexes them into a searchable archive with citations that click back to the source line. The same pattern works for anything you have a lot of text about: a medical textbook collection, a guitar theory library, a field's research papers, a car's service manuals, your company's internal wiki. Whatever corpus you give it becomes a searchable, talkable version of exactly what you have.

Grounding for AI agents

For programmers, lilbee plugs into whatever AI agent you already use (via MCP). Feed it your project's docs, your dependency source, the vendor SDK reference, your design notes, and the agent stops making up function names. It reads the actual code it's about to call, cites the file and line, and tells you when the answer isn't in the corpus instead of guessing. That matters: a lot of AI today produces confident-sounding guesses and charges per token for them. lilbee is built the other way. Answers should come from sources you can check, and the system should be willing to say it doesn't know.

Offline copies of websites

Web crawling paired with local search and chat takes one command. Install the crawler extra, point lilbee at a docs site, a wiki, or a vendor's API reference, and the pages get fetched, converted to markdown, and indexed. From then on you can search or chat offline with that image of the site, even if it changes or goes down.

How it's built

Under the hood lilbee stands on established open-source projects: Kreuzberg handles document parsing, LanceDB is the embedded search layer, llama-cpp runs models locally, crawl4ai and Playwright crawl the web, and Textual draws the terminal. The architectural bet is that everything stays embedded in one process. Most systems in this shape deploy a vector database and a model server separately, usually reaching for a cloud-hosted search service- which moves your data onto someone else's servers. lilbee performs the entire process locally. Pull the module to your laptop, pull your models, and curate your documents. A complete local search-and-chat stack with nothing to deploy.

Documents, code, and scanned images

Document and code processing get treated as a first-class problem. Most retrieval libraries throw your files at a PDF extractor and call it done. lilbee splits the work along the grain of what's being indexed: prose and structured documents (90+ formats across PDFs, Office files, ebooks, HTML, and more) go through Kreuzberg's Rust-based extraction pipeline with heading-aware chunking, so each chunk keeps its section context. Code goes through tree-sitter's AST-aware splitter across 150+ languages, so chunks map to real functions, classes, and modules instead of arbitrary line ranges. Retrieval returns things that make sense on their own, not fragments that cut through an argument or a function signature.

Scanned PDFs and photographed notes go through an OCR pipeline with a choice of backends: Tesseract, a local GGUF vision model via llama-cpp's mtmd backend (which preserves tables and layout as markdown), or a remote vision model through the SDK backend.

Pick and tune your models

Chat, embedding, vision, and reranking models are installed and switched from inside the terminal: browse the catalog, pull a GGUF, pick a role. Retrieval and generation are deeply tunable. You can make chunks smaller for finer-grained matches, make search stricter to filter out loose results, skip automatic query rewriting for faster responses, turn on a second-pass re-scorer for precision over the top results, or lean more on topic relationships when your corpus has lots of interconnected ideas. All editable from the TUI, environment variables, or a project-local config file, with sensible defaults out of the box.

Local-first, frontier-capable

lilbee is built as a local-first tool, but can also connect to cloud-hosted models. The TUI supports API keys and shows a persistent warning whenever a cloud-hosted model is active so it's clear when chunks are leaving the machine. When a local model isn't enough you can start to access popular frontier models with the commands pip install --pre 'lilbee[litellm]' or `uv tool install --prerelease=allow 'lilbee[litellm]''

TUI

lilbee with no args (or lilbee chat) launches a full Textual terminal app. Chat streams replies with clickable citations. A Task Center tracks every background job (sync, crawl, wiki build, model pull) and lets you cancel them with /cancel. Other screens cover the model catalog (/models), settings (/settings), first-time setup wizard (/setup), and the auto-built wiki (/wiki). Tab completion works for slash commands, file paths, model names, setting keys, and themes.

See Previews for a visual and the slash-command reference for the full list.

Hardware requirements

Standalone mode runs entirely on your machine. No cloud required.

Supported platforms

Resources

Popular frontier models are optional; install with pip install --pre 'lilbee[litellm]' or uv tool install --prerelease=allow 'lilbee[litellm]'.

Install

Prerequisites

• Python 3.11, 3.12, 3.13, or 3.14
• Optional (for scanned PDF / image OCR): Tesseract (brew install tesseract / apt install tesseract-ocr) or a GGUF vision model (see vision OCR)

No external services needed. lilbee downloads and runs GGUF models locally via llama-cpp-python.

Default install (recommended for almost everyone)

The default wheel ships with runtime CPU dispatch (works on every x86_64 CPU from Sandy Bridge / 2011 forward) and GPU acceleration via Vulkan on Linux/Windows (covers NVIDIA / AMD / Intel Arc) or Metal on macOS arm64. If you have any modern GPU, this gets you GPU-accelerated inference with zero opt-in.

pip:

pip install --pre lilbee

uv:

uv tool install --prerelease=allow lilbee

Verify the install works on your hardware:

lilbee self-check

This downloads a tiny model (~90 MB), runs an inference, and an embedding. Exits 0 with SELF-CHECK PASSED on success.

NVIDIA users wanting CUDA-native (5–15% faster than Vulkan)

The default wheel already uses your NVIDIA GPU through Vulkan. You only need a CUDA wheel if you want the absolute last bit of performance out of CUDA-native kernels.

CUDA wheels live on a per-CUDA-version index (because each is linked against a specific CUDA runtime). Pick the index that matches your installed CUDA toolkit:

# CUDA 12.4 — pip
pip install --pre lilbee --extra-index-url https://tobocop2.github.io/lilbee/cu124/

# CUDA 12.4 — uv tool
uv tool install --prerelease=allow lilbee \
  --extra-index-url https://tobocop2.github.io/lilbee/cu124/

# CUDA 12.5 — pip
pip install --pre lilbee --extra-index-url https://tobocop2.github.io/lilbee/cu125/

# CUDA 12.5 — uv tool
uv tool install --prerelease=allow lilbee \
  --extra-index-url https://tobocop2.github.io/lilbee/cu125/

Don't know your CUDA version? nvidia-smi (look at the top-right corner). NVIDIA driver 555+ supports CUDA 12.5; older drivers may need cu124.

Intel Mac

Intel Mac wheels ship from a separate index because they're cross-compiled and the lane is best-effort:

pip install --pre lilbee --extra-index-url https://tobocop2.github.io/lilbee/cpu/

If a wheel isn't available for your Python version, pip falls back to the sdist and builds locally (you'll need a working Xcode toolchain).

Pre-built executables (no Python required)

Download a single binary that bundles its own Python runtime — no pip needed:

Make executable and run:

chmod +x lilbee-linux-x86_64
./lilbee-linux-x86_64 self-check

Always points at the latest pre-release.

Optional extras

lilbee works out of the box. Extras unlock additional capabilities. Both pip and uv tool install syntax shown:

Install multiple at once:

pip install --pre 'lilbee[graph,crawler,litellm]'
# or
uv tool install --prerelease=allow 'lilbee[graph,crawler,litellm]'

Combine with --extra-index-url for CUDA:

uv tool install --prerelease=allow 'lilbee[graph,crawler]' \
  --extra-index-url https://tobocop2.github.io/lilbee/cu125/

See the full guide on optional extras for configuration and details.

Upgrading to the latest pre-release

# pip
pip install --upgrade --pre lilbee

# uv (force reinstall)
uv tool install --reinstall --prerelease=allow lilbee

Development (run from source)

git clone https://github.com/tobocop2/lilbee && cd lilbee
uv sync
uv run lilbee

Agent integration

lilbee serves as a retrieval backend for AI coding agents via two entry points: an MCP server (lilbee mcp) and a JSON CLI (lilbee --json ...). MCP exposes search, document lifecycle, crawling, model management, and the full wiki surface as tools; search takes a scope argument so agents can target documents, wiki pages, or both.

See docs/agent-integration.md for MCP client configuration, the full tool reference, and JSON CLI examples.

HTTP Server

lilbee serve starts a REST API that any tool or GUI can hit. It covers search (with SSE streaming), document lifecycle, crawling, model management, configuration, and vault-aware source retrieval for GUI clients. Interactive API docs live at /schema/redoc when the server is running.

See the API reference for the full OpenAPI schema and the usage guide for serve options.

An Obsidian plugin pairs with lilbee for users who want a GUI alongside the terminal. It runs lilbee serve in the background, so there's no separate service to manage, and every citation in chat or wiki opens a Source Preview that scrolls to the exact passage in the original document — useful when seeing the source matters more than just being told where it is. Install via BRAT; see the plugin README for the four-step setup.

Interactive chat

Running lilbee or lilbee chat enters the TUI. Type / to see the full slash-command list inline, or check the slash-command reference in the usage guide. Slash commands and paths tab-complete; background jobs appear in the Task Center and are cancellable with /cancel.

Supported formats

Text extraction powered by Kreuzberg, code chunking by tree-sitter. Structured formats (XML, JSON, CSV) get embedding-friendly preprocessing. This list is not exhaustive; Kreuzberg supports additional formats beyond what's listed here.

See the usage guide for OCR setup and model benchmarks.

Experimental

Two opt-in features that work but are still finding their final shape. Generation quality and retrieval behavior depend on corpus, models, and knobs; expect to iterate. Feedback is welcome.

Wiki

lilbee analyzes the documents you've indexed and writes a wiki about them. Pages compound across sources instead of being one-per-document, so concepts and entities that show up repeatedly get their own page with citations from every source that mentions them. Pages live under $LILBEE_DATA/wiki/, grouped into concepts/, entities/, and a drafts/ queue when confidence is low. An index.md tracks them all and log.md records every build, ingest, and prune.

Every section is citation-verified against the source chunks and scored for embedding faithfulness before publish. Plain-text concept slugs inside page bodies are rewritten to [[wiki link]] form so graph-style markdown viewers can render the connections. Some pages will land in drafts/ for human review rather than publish direct.

See the Wiki section of the usage guide for the full command list and configuration.

Semantic chunking

A semantic-chunking mode is available as an opt-in alternative to the default fixed-size chunker. It uses embedding similarity to find topic boundaries, so each chunk is one coherent thought instead of a fragment that cuts through an argument. The benefit shows up on prose-heavy corpora like novels, essays, long-form research papers, or interview transcripts. The trade-off is roughly 9x more embedding calls during indexing.

See the Semantic chunking section of the usage guide for trade-offs and how to enable it.

License

Elastic License 2.0 (ELv2). See LICENSE.