Jaroslav Beneš
17f3fa1946
A long conversation eventually just stops working. Compaction summarises the earlier turns and sends the summary in their place. The messages are kept. They stay in the transcript behind a collapsed divider and simply stop being part of the request, which is what makes the button safe to press and automatic compaction safe to have at all: a summary that came out badly is a bad turn, not a lost conversation. Stored on the Chat, not as a synthetic Message. A synthetic row needs a role -- `system` breaks the one-system-message rule the moment build_messages emits it beside the harness, and user/assistant makes it a turn people can edit, regenerate from and copy, indistinguishable from a real one in all four places a bubble is rendered. Worse, "editing rewinds, it does not branch" would silently delete it and leave no marker that compaction had happened at all. The summary goes out as a user turn and an assistant turn, not one. A leading assistant breaks templates requiring the first non-system message to be user; a lone leading user produces user, user whenever the kept history starts on a user turn -- which it always does, because the cutoff lands on a finished reply. compacted_through_id is a plain id rather than a foreign key: migrations.py compiles only the column type, so a REFERENCES clause would exist on a fresh database and not on an upgraded one, and a constraint half the fleet has is worse than none. cutoff_message validates it on every read instead, and a rewind past the boundary clears it. Compacting again summarises only the delta, with the previous summary supplied to be subsumed. Re-summarising the whole chat each time grows quadratically and eventually exceeds the window it is protecting. Automatically at the top of _run, not in post_message: that route's contract is to return immediately and leave the slow part to a resumable connection, and it also means build_request is called once, after compaction, with no second assembly path. The trigger is the last reply's recorded usage plus an estimate of the new turn -- retrospective because true prompt_tokens are only knowable after a response, plus the delta because otherwise fifty thousand characters pasted into the composer overflow a window that read 90% last turn. It never fires when the context length is unknown. It does fire on estimated counts, which is safe here precisely because nothing is lost. _maybe_compact never raises: a failure logs and sends the uncompacted request. A `status` event says "Summarising earlier messages…" in the meantime, because a silent multi-second pause before the first token is what a hang looks like. The wording is three fragments under Admin - Prompts. Clearing task.compact turns compaction off entirely. Also adds compaction.moment(): SQLite does not store the offset, so a row loaded from disk is naive while one in the session's identity map keeps its tzinfo, and comparing the two raises. Every comparison here is between exactly those. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
A self-hosted web UI for your language models, written in Python.
Talks to anything that speaks the OpenAI API. Themed after Middle-earth.
Lembas is the Elvish waybread — one bite sustains a traveller for a day's march. The capitals hide what it runs on: LLeMbas.
Why this exists
Most self-hosted LLM front-ends are large JavaScript applications with a Python
API bolted underneath. LLeMbas is the other way round: server-rendered
Python, with htmx and a little Alpine for interactivity. There is no
package.json, no bundler, no build step, and nothing is fetched from a CDN at
runtime. Clone it, pip install -e ., run it.
Features
Working now
- Chats — streaming replies, Markdown with server-side syntax highlighting, copy and regenerate, automatic chat titles. Chats are created when you send the first message, so an abandoned one never clutters the sidebar
- System prompts — instance-wide, per-model and per-chat, with the most specific winning outright
- Reasoning display — thinking streams into its own collapsible block (closed by default), labelled with how long it took, and is never replayed as context
- Live Markdown — formatting appears as the model writes, not at the end
- Stop and rewind — cut a reply short and keep what arrived, or edit an earlier message and run the conversation on from there
- Replies keep running in the background — navigate away, open another chat, close the tab; a green dot and a notification tell you when it lands
- Attachments — drag, paste or pick images, PDFs and text files. Images are downscaled and sent to vision models; PDF and text content is extracted and put in the prompt
- Folders — arbitrarily nested, delete a folder without losing the chats inside it
- Web search — offered to the model as a tool it calls when a question needs it. DuckDuckGo out of the box (no account, no key), or point it at your own SearXNG, or Firecrawl. The sources stay in the transcript
- Speech in and out — dictate a message and have replies read aloud, against any OpenAI-compatible audio endpoint (whisper.cpp, Speaches, Kokoro…). Each person picks their own voice
- A library — four places a model can reach for. Knowledge: documents, images and web pages you collect, grouped into named bases so a chat can be pointed at just the right one, searched before the web. Notes: longer things it writes down and finds again later. Memory: short facts about you, in front of it on every turn. Skills: saved procedures it can follow, and write. All of it visible and editable by you, and shareable with a group or a person, read-only
- Installable — add it to a phone home screen or a desktop launcher and it runs in its own window
- OpenAI connections — point at OpenAI, LM Studio, vLLM, llama.cpp, llama-swap, Ollama or OpenRouter; models are discovered and cached
- Model settings — searchable, filterable list with a page per model: ordering, pinned models, an instance default and a per-user default, custom names, descriptions and images. Scales to hundreds of models
- Users, groups & permissions — per-group grants that union rather than override, and model access restricted to chosen groups
- Accounts — first account becomes the administrator, argon2 password hashing, revocable server-side sessions, self-service password change, admin-managed accounts
- Admin settings — open or close registration from the UI, stored in the database and effective immediately
- Two themes — Moria (dark) and Shire (light), switchable per user
Planned
Custom tools and MCP servers · agentic execution (local and over SSH) · image generation · OCR for scanned PDFs · semantic search in the library.
See PLAN.md for what is built, what is not, and why.
Quick start
git clone https://git.houmeres.sk/Houmeres/LLeMbas.git
cd LLeMbas
python -m venv .venv && . .venv/bin/activate
pip install -e ".[dev,search]" # `search` adds DuckDuckGo; drop it if unwanted
cp .env.example .env
lembas secret-key # paste the result into LEMBAS_SECRET_KEY
lembas serve # http://127.0.0.1:8080
Open the address and create the first account — it becomes the administrator.
Then go to Admin → Connections and add an endpoint. For a local runner that
is usually http://localhost:1234/v1 with no API key. Press Test & refresh
and its models appear in the chat model picker.
The vendored browser libraries (htmx, Alpine) are committed, so no network access is needed to run. To re-fetch or bump them:
python scripts/fetch_vendor.py --update.
Web search
Admin → Web search. DuckDuckGo needs nothing beyond the search extra
above. SearXNG needs its JSON format enabled — add - json under
search.formats in its settings.yml, or every search fails. Firecrawl needs
an API key.
Search is offered to the model as a tool, so it decides when a question needs looking up. It is only offered to models marked tools under Admin → Models: an endpoint without tool support rejects the whole request rather than ignoring the extra field, so the flag is a real switch and not a hint.
Audio
Admin → Audio. Two endpoints, because they are usually two servers:
| Speaks | Example | |
|---|---|---|
| Dictation | POST /v1/audio/transcriptions |
whisper.cpp's whisper-server, Speaches, faster-whisper-server |
| Read aloud | POST /v1/audio/speech |
Kokoro-FastAPI, OpenAI |
If the speech endpoint also answers GET /v1/audio/voices the voice list is
read from it, and each person can pick their own under Settings → Audio.
Recorded audio is passed straight through and never written to disk.
The microphone needs HTTPS or localhost. Browsers do not grant it over plain HTTP, so a LAN install without TLS will not offer dictation.
The library
Sidebar → Library, and Settings → Memory. Nothing is on by default for a
model: give it the tools it should have under Admin → Models, where
tools decides whether a tool list may be sent at all and the built-in tools are
chosen one by one.
Knowledge is organised into bases — one per subject, project or client. A chat with no base attached searches everything you have; tick some in the chat's settings panel and it searches only those. Sharing happens at the base: share it and everything in it comes too, read-only.
Search is SQLite's FTS5 — keyword matching with BM25 ranking, no embedding service to run and nothing that stops working offline. It will not match a paraphrase, so a line of description on a document is worth writing.
Saving a link makes your server fetch a URL. Addresses on your own machine and network are refused unless an administrator opts in under Admin → Web search, because the address can come from a model and the server can reach things your browser cannot.
Installing as an app
Open it in a browser and use Install (Chromium) or Share → Add to Home Screen (iOS). This also needs HTTPS or localhost — service workers are unavailable over plain HTTP, and without one there is nothing to install.
There is no offline mode beyond a page saying so. Everything is rendered by your server, so a cached conversation would be a snapshot that silently went stale.
Configuration
All variables are prefixed LEMBAS_ and can live in .env. See
.env.example for the annotated list.
| Variable | Default | Purpose |
|---|---|---|
LEMBAS_SECRET_KEY |
generated | Signs sessions and encrypts stored API keys. Set this. A generated key changes every restart, signing everyone out and making stored API keys unreadable. |
LEMBAS_DATA_DIR |
./data |
SQLite database and uploads. |
LEMBAS_HOST / LEMBAS_PORT |
127.0.0.1 / 8080 |
Bind address. |
LEMBAS_ALLOW_SIGNUP |
true |
Whether new users may register themselves — the initial value only. Once set under Admin → General the stored setting wins. The first account is always an admin regardless. |
LEMBAS_DEFAULT_THEME |
moria |
moria (dark) or shire (light). |
LEMBAS_SESSION_TTL |
2592000 |
Session lifetime in seconds. |
LEMBAS_REQUEST_TIMEOUT |
300 |
Seconds to wait on an upstream model. |
Commands
lembas serve # run the server
lembas info # where data lives, what is configured
lembas secret-key # generate a value for LEMBAS_SECRET_KEY
lembas create-admin # create or promote an administrator
How it fits together
Browser ──form POST──▶ FastAPI ──▶ SQLite
▲ │
│ └──httpx──▶ any OpenAI-compatible endpoint
└──── server-sent events ◀───────────────┘ (streamed reply)
Sending a message stores the turn and returns two HTML fragments: the user's
bubble and an empty assistant bubble carrying an sse-connect. That opens a
server-sent event stream which appends tokens as they arrive, then replaces the
whole bubble with the finished, Markdown-rendered version. Rendering and
highlighting happen in Python, so the streamed and final views cannot disagree.
src/lembas/
api/ routes: auth, chats, folders, admin, pages
db/models/ SQLAlchemy schema
security/ password hashing, sessions
services/ llm client, chat orchestration, markdown, crypto, sse
web/ Jinja templates and static assets
assets/ SVG artwork masters
scripts/ artwork generator, vendored-JS fetcher
deploy/ systemd unit and nginx vhost for a real install
Development
pytest # test suite
ruff check . # lint
python scripts/build_artwork.py # regenerate the SVG artwork
python scripts/fetch_vendor.py # verify vendored JS against the lockfile
There is no Alembic. The schema is SQLite-only and synchronised at startup:
missing tables and missing columns are added automatically, so adding a field to
a model needs nothing but a restart. Renames, drops and retypes are still manual
— see CLAUDE.md.
Artwork
The logo, favicon and banner are original vector work, generated by
scripts/build_artwork.py so the mallorn leaf stays
identical across every size it appears at. The wordmark is
Source Serif 4 (SIL OFL 1.1)
converted to outlines — a README banner cannot load a webfont, and <text>
would render in whatever serif the reader happens to have.
Licence
A note on the theme
This is an independent hobby project, themed as an affectionate nod to J.R.R. Tolkien's world. It is not affiliated with, endorsed by, or connected to the Tolkien Estate, Middle-earth Enterprises, or any related rights holder. All artwork here is original.