Lilbee

What is Lilbee about?

Lilbee is a self‑contained RAG platform that indexes personal documents, codebases, and web content, then lets a local LLM answer questions or hold a chat using those sources. All components (vector store, model runtime, crawler, wiki builder) run in a single process, reachable via a Textual TUI, a REST API, an MCP server, or directly as a Python library.

How to use Lilbee?

1. Install – pip install --pre lilbee (or use Homebrew, Docker, etc.).
2. Initialize – run lilbee self-check to verify hardware and download a tiny model.
3. Add data – point Lilbee at a folder (lilbee add /path/to/docs) or use the built‑in crawler (lilbee crawl https://example.com).
4. Sync – lilbee sync builds embeddings with LanceDB and creates a searchable vault.
5. Chat / Search – launch the TUI with lilbee (or lilbee chat). Type queries; switch between Search (retrieval‑augmented) and Chat (pure generation) with F3.
6. Advanced access – start the MCP server (lilbee mcp) or the REST API (lilbee serve) for integration with other agents or GUIs.

Key features of Lilbee

• Terminal‑first UI built with Textual (TUI, slash‑commands, clickable citations).
• Hybrid search: vector similarity via LanceDB + optional BM25 lexical fallback.
• OCR pipeline: Tesseract or local GGUF vision models for scanned PDFs and images.
• Auto‑generated wiki with concept/entity pages and citation‑verified drafts.
• Model catalog: browse, pull, and switch local GGUF models for chat, embedding, rerank, and vision.
• Web crawling (optional extra) using Playwright & crawl4ai, converting pages to markdown.
• Multi‑role inference: separate subprocesses for chat, embedding, rerank, and vision keep the UI responsive.
• Offline‑first operation; optional frontier/cloud models via litellm extra.
• MCP server & REST API for programmatic retrieval, document management, and tool integration.

Use cases of Lilbee

• Personal knowledge base: index PDFs, notes, ebooks, and retrieve answers with citations.
• Developer assistance: ground code‑generation agents in the actual project source tree.
• Offline documentation browsing: crawl vendor docs once, then search/chat without internet.
• Research workflow: index a collection of papers and ask natural‑language questions with source links.
• Education: create a searchable wiki from textbooks and lecture slides.
• Small‑team internal docs: self‑host a private RAG system with full data privacy.

FAQ from Lilbee

Q: Do I need an internet connection? A: No. All indexing, embedding, and inference run locally. Internet is only required to download models or optional frontier services.

Q: Which hardware is required? A: Minimum 8 GB RAM and a 2013‑or‑newer 64‑bit CPU. GPU acceleration works via Vulkan (Linux/Windows) or Metal (macOS) and optional CUDA wheels for NVIDIA.

Q: Can I use cloud LLMs? A: Yes, by installing the litellm extra. The UI shows a persistent warning whenever a cloud model is active.

Q: How are code files chunked? A: Tree‑sitter parses over 150 languages, creating AST‑aware chunks that correspond to functions, classes, or modules.

Q: Is there a way to extend the model catalog? A: The catalog automatically lists locally available GGUF files and, when API keys are configured, frontier models from OpenAI, Anthropic, Gemini, etc.

Q: How do I integrate with other AI agents? A: Run lilbee mcp to expose MCP endpoints (search, document CRUD, wiki, model management) that agents can call as tools.

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 built-in model 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]  Embed [Nomic v1.5]   [Search|Chat]     │ │
 │ └───────────────────────────────────────────────────────────┘ │
 │ 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. The Local sub-tab covers anything you can run on this machine (native GGUF and any locally-running SDK provider). The Frontier sub-tab appears when at least one cloud-provider API key is configured and lists what each provider exposes, grouped by Anthropic, Gemini, OpenAI, and so on. * indicates the developer's recommendation for each role.

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

Picking a row on Frontier sets that model as the active chat model. Frontier listings come straight from the SDK backend's view of each provider; lilbee does not curate or filter the list.

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 model, 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.

Frontier-capable, no cloud round-trips by default

lilbee runs entirely on your machine, 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.

The model bar above the prompt has searchable pickers for the active chat and embedding models (click or press Enter on the button to open a modal with a search box and a virtualized list; type to filter, Enter to pick, Escape to cancel). Next to the pickers is a Search / Chat toggle (also bound to F3). In Search mode every prompt runs through document retrieval; in Chat mode retrieval is skipped and the model answers directly. Search mode falls through to a chat answer when nothing relevant is indexed, and shows a one-time toast when that happens. The toggle is forced to Chat and disabled when no embedding model is configured.

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

Platform	Minimum	Recommended
Linux x86_64	A 64-bit Intel or AMD CPU from 2013 or newer — Intel Core i3/i5/i7 4th-gen (Haswell), Intel Xeon E3-12xx v3 / E5-26xx v3, AMD FX-95xx (Steamroller) or any AMD Zen-based chip. Anything corresponding to the x86-64-v3 microarchitecture level.	A modern Intel Core / Xeon / AMD Ryzen / EPYC + an NVIDIA, AMD, or Intel Arc GPU
macOS arm64	Any Apple Silicon Mac (M1 or newer) running macOS 11+	M-series Pro / Max / Ultra
Windows x86_64	A 64-bit Intel or AMD CPU from 2013 or newer (same generations as Linux above), Windows 10/11	Modern desktop / workstation CPU + GPU
Linux ARM64	ARMv8 (NEON-capable) — Raspberry Pi 4+, AWS Graviton, Ampere Altra, etc.	Modern ARM server with 16+ GB RAM

Resources

Resource	Minimum	Recommended
RAM	8 GB	16 to 32 GB if you load several local models at once (chat + embed + rerank + vision); the actual footprint scales with the size and quantization you pick
GPU / Accelerator	none required (CPU-only inference works)	Apple Silicon (Metal) · any NVIDIA / AMD / Intel Arc GPU (Vulkan) · NVIDIA GPU + matching CUDA toolkit (opt-in CUDA-native wheels, see Install)
Disk	2 GB (models + data)	10+ GB if you load multiple models

Each active inference role (chat, embed, rerank, vision) runs in its own subprocess to keep the TUI responsive; the resident memory you'll need follows the size of the models you keep warm. 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 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.

Homebrew (macOS arm64, Linux x86_64)

If you'd rather not install Python, the prebuilt binary is available on a Homebrew tap. It bundles its own Python interpreter and llama-cpp backend, the same artifact that ships on the GitHub Release page.

brew tap tobocop2/lilbee
brew install lilbee

The macOS binary is unsigned (Apple's developer certificate costs $99/year). The formula clears the com.apple.quarantine extended attribute automatically during install, so the first launch is not blocked by Gatekeeper. If macOS still blocks it, open System Settings → Privacy & Security and click Allow Anyway.

Arch Linux (AUR)

Available as lilbee for paru / yay / pacaur / any AUR helper:

paru -S lilbee

Wraps the Linux x86_64 release binary, no compilation needed.

Docker

docker run --rm -v lilbee-data:/home/lilbee/data ghcr.io/tobocop2/lilbee:latest --help

Image is published to GitHub Container Registry on every release; tagged with both the version (0.6.66b456) and latest. The LILBEE_DATA_DIR is /home/lilbee/data inside the container, so mount a volume there to persist models, embeddings, and config.

Nix (NixOS, nix-darwin, or any host with nix)

nix run github:tobocop2/lilbee

Wraps the same release binary as Homebrew / AUR / Docker. On Linux the flake bundles glibc, libgomp, and vulkan-loader so it runs on bare NixOS without libvulkan1 on the host.

Linux runtime requirements

The Linux x86_64 wheel links against the Vulkan loader at runtime so it can fall back from GPU to CPU on a single binary. Most desktop distros (Ubuntu 22.04+, Pop!_OS, Mint) ship libvulkan1 by default. Bare Arch / Fedora / Alpine images do not, and lilbee self-check will fail with cannot open shared object file: libvulkan.so.1. Install the loader once.

# Arch / Manjaro
sudo pacman -S vulkan-icd-loader

# Fedora / RHEL
sudo dnf install vulkan-loader

# Debian / Ubuntu (only if missing)
sudo apt-get install libvulkan1

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:

Platform	Download
Linux x86_64	lilbee-linux-x86_64
macOS arm64	lilbee-macos-arm64
Windows x86_64	lilbee-windows-x86_64.exe

Make executable and run:

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

Always points at the latest pre-release.

The Linux binary is built on manylinux_2_28 and requires glibc 2.28 or newer, which covers every currently-supported distribution: Fedora 28+, RHEL/AlmaLinux/Rocky 8+, Debian 10+, Ubuntu 18.10+, Amazon Linux 2023, and current Arch. For older systems (CentOS 7, Amazon Linux 2, RHEL 7), use uv tool install lilbee or the Docker image.

Optional extras

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

Extra	What it adds
Web crawling — pip install --pre 'lilbee[crawler]' / uv tool install --prerelease=allow 'lilbee[crawler]'	Index websites alongside local files. Recursive crawling with Playwright, live progress, cancel, hash-based change detection, SSRF protection, rate limits.
Ollama and frontier models — pip install --pre 'lilbee[litellm]' / uv tool install --prerelease=allow 'lilbee[litellm]'	Keep compatibility with existing Ollama setups, or use a popular frontier model (OpenAI, Anthropic, Gemini, etc.) for chat, vision, or embeddings while keeping other roles local. You provide the API key. Chunks sent to the provider leave your machine, and the TUI shows a persistent warning while a cloud model is active.
Concept graph — pip install --pre 'lilbee[graph]' / uv tool install --prerelease=allow 'lilbee[graph]'	Topic clustering and search boosting. Extracts concepts from your documents and uses their relationships to find results pure text matching misses. Zero extra LLM calls.

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.

Format	Extensions	Requires
PDF	.pdf	none
Scanned PDF	.pdf (no extractable text)	Tesseract (auto, plain text), or a GGUF vision model via the native mtmd backend (recommended, preserves tables, headings, and layout as markdown)
Office	.docx, .xlsx, .pptx	none
eBook	.epub	none
Images (OCR)	.png, .jpg, .jpeg, .tiff, .bmp, .webp	Tesseract
Data	.csv, .tsv	none
Structured	.xml, .json, .jsonl, .yaml, .yml	none
Code	.py, .js, .ts, .go, .rs, .java and 150+ more via tree-sitter (AST-aware chunking)	none

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.