Chipper ships two lockfiles, no install command, and a launcher keyed to the GPU profile
✨ AI interface for tinkerers (Ollama, Haystack RAG, Python)
At a glance
- What is it?
- Chipper is a containerised Python RAG service that also works as a proxy in front of Ollama, built by one person as a private tool for a book project. Its mechanics are more interesting than its documentation: the install section contains two links and no commands, the launcher picks between container, host and localhost Ollama URLs from the GPU profile, and the README's own roadmap still lists an automated unit testing framework as upcoming.
- Who is it for?
- Chipper rewards a reader who opens setup.py instead of the documentation site, because the launcher, the three Ollama URL constants and the reachability probe are where the real behaviour lives, and the file states plainly that it is a personal project not designed for commercial or production use.
- Can I use it commercially?
- Yes. MIT is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository last received commits 180 days ago.
- What is it written in?
- Mainly Python, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on November 15, 2026, and from our analysis. They are not legal advice.
Star history
- TilmanGriesel/chipper487 ★ · +4
| Repository | 2026-09-11 | 2026-11-15 |
|---|---|---|
| TilmanGriesel/chipper | 483 | 487 |
4 Stars in the last 65 days, measured from the daily GitHub snapshot. Compare with other projects
Editorial analysis
The install section contains two links and no commands
The Installation and Setup heading in the README holds no shell commands, no compose file and no configuration snippet. It holds one link to a quickstart anchor on the project website and one link to the project website itself. Everything a new user needs in order to start the service therefore lives outside the repository, in a documentation site built with VitePress, and the one example directory the tree does contain is examples/docker_compose/. A reader who clones the repository expecting a compose file at the root will find a docker/ directory, a services/ directory, a run file and a run.sh file instead, with nothing in the README saying which one starts the stack. The overview list at the top of the README offers seven anchors, and one of them, Friends of Chipper, has no visible section of its own in the part of the file that follows, so the navigation is slightly wider than the content behind it. The one line description field opens with a decorative character and calls the project an AI interface for tinkerers, naming Ollama, Haystack RAG and Python. The README explains where that came from: it began as a personal tool to help the author's girlfriend with her book, using local retrieval and language models to explore characters and creative ideas while keeping her work private and off cloud services such as ChatGPT. A few handy scripts became a containerised service, and the philosophy section frames the goal as tools approachable for beginners and still useful to experts, offered to educators as a framework for teaching and to companies as a playground.
setup.py picks the Ollama URL from the platform and the GPU profile
The launcher defines three default Ollama endpoints and chooses between them. The internal default is http://ollama:11434, which is how a container reaches another container on the same network. The external default is http://host.docker.internal:11434, with a comment in the source that says outright that this is not linux compatible and links to a question about the Linux equivalent. The third default is http://localhost:11434, for the case where Ollama is on the same machine. The selection function reads the platform name and the platform release, logs both along with the GPU profile it was handed, then treats an external URL as required when the platform is Darwin, or when the profile is cpu, or when the profile is amd-linux. Any other combination, meaning a non Darwin host running a profile that is neither cpu nor amd-linux, falls through to a branch that the visible part of the file never reaches. All three constants carry the same port, and the reachability check uses that same number as its fallback when a URL carries none, so the port appears in four places in this one file and is configurable in none of them.
The reachability probe only opens a TCP socket
Before choosing an endpoint the launcher calls a check that parses the URL, falls back to port 11434 when the URL carries no port, opens an AF_INET stream socket, sets a five second timeout and calls connect_ex. It reports the endpoint as available when the result is zero. That is a TCP connect, so anything listening on the port satisfies it: the probe never sends an Ollama request, never asks which models are pulled, and never distinguishes an Ollama instance from another process holding the same port. The socket family is fixed to AF_INET, so an endpoint that only resolves to an IPv6 address is not covered by this path at all. The launcher also defines EXAMPLE_API_KEY with the literal value EXAMPLE_API_KEY, and leaves the shared key and the external URL as None until something fills them in, which is a reminder that the defaults here are scaffolding for a manual edit rather than a finished configuration. The module imports secrets, shutil and subprocess next to the socket code, so whatever the launcher does beyond opening a port involves key material, file movement and child processes, though the visible part of the file stops inside the external Ollama branch before any of that is resolved.
pyproject.toml configures two formatters and nothing else
The Python metadata is not in pyproject.toml. That file holds a black section with a line length of 88 and an exclude pattern covering the git directory, virtual environments, bytecode caches, build and dist, and an isort section whose profile is black. There is no build system table, no project table, no dependency list and no entry point anywhere in it. The packaging metadata lives in setup.py, which is a Python module rather than a declarative file and which imports argparse, os, platform, secrets, shutil, socket and subprocess alongside pathlib and urlparse. Around it the root carries four more tool configurations: .flake8, .editorconfig, .gitattributes and .pre-commit-config.yaml. So a reader looking for the dependency list of a service that talks to Ollama, Elasticsearch and Hugging Face finds it in a script whose job also includes deciding container networking, and looking for it in pyproject.toml finds formatting rules instead. The root also holds a tools/ directory, a docs/ directory that the JavaScript scripts build, and a services/ directory, so the launcher, the container definitions and the documentation site each live in their own place and none of the three is mentioned in the setup section.
Two lockfiles guard a toolchain that only builds documentation
The root carries both package-lock.json and yarn.lock, and the JavaScript manifest pins its package manager to yarn 1.22.22 with a long integrity hash. Keeping both is normally the sign of a migration in progress. What the JavaScript side actually does here is small: four devDependencies, autoprefixer, postcss, tailwindcss and vitepress, and three scripts, docs:dev, docs:build and docs:preview, all pointed at the docs directory. So the entire Node toolchain builds the documentation site and the Tailwind styles for the web interface, and the only front-end code in the manifest is the Tailwind toolchain. The web interface itself is described as vanilla JavaScript that works offline without a framework, with a /help command for switching models and updating the embeddings index, and automatic syntax highlighting for popular languages in its code output view. Alongside the two lockfiles the root also holds two launchers, a file named run and a run.sh, plus update_requirements.sh for the Python side.
The roadmap sits inside the demos section and unit tests are still a checkbox
The Project Roadmap is not a top level section. It appears as a third level heading below the Demos section, with its completed milestones and upcoming features nested under it at the fourth and fifth levels, so the plan for the project is filed under a heading about screenshots and screenshots of code output. The completed list covers the basics, the command line interface, the web interface, containerisation, better mobile support, linting and formatting, publishing an image to Docker Hub, client side text to speech, bearer token authentication, mirroring the Ollama chat API, two Haystack chat generators and distributed processing. The upcoming list is short and includes an automated unit testing framework, expanded client side model settings, an initial set of three tutorials, smarter document splitting and embedding, and the React application. An unchecked test framework next to a checked linting milestone is the line worth reading twice. The tutorials item is scoped as an initial set of three for new users, and smarter document splitting and embedding sits beside it as another open box, which reads oddly against a feature list that already claims document chunking as shipped.
Four shields with no text and a TODO about one of them
The header opens with an empty centred paragraph, then a commented out link to the project website carrying a TODO that asks why this shield does not work anymore. Four visible anchors follow, and all four have no link text at all: one to the actions page, one to the stargazers page, one to the issues page and one to a Docker Hub repository. The Docker Hub target is worth noting on its own, because its namespace is griesel rather than tilmangriesel, so the container image is published under a different owner prefix than the GitHub account. That same name split shows up in the setup script, whose default API key constant is a placeholder rather than a secret, and in the release history, where the tags are v2.5.0 and v2.4.0 but the third is v.2.3.0 with a dot in front of the number, while the release titles switch between a bare tag and a Chipper prefixed name. The topic list attached to the repository is empty, so none of the framework names are indexed anywhere, and the acknowledgments name Haystack, Ollama, Hugging Face, Elastic and Elasticvue before a final entry whose line stops right after an opening square bracket.
The proxy adds authentication, and the file rules out production
The feature list describes Chipper as sitting between an Ollama client and an Ollama instance, naming Enchanted and Open WebUI as the clients it aims at, and the third party section says every such client gains server side knowledge base embeddings, server side model selection, query parameters and system prompt overrides. That makes Chipper a middleware in front of a model server rather than only a search tool, which is why the API security bullet exists: it promises API key based authentication and what the file spells Baerer token service authentication, a typo for Bearer. Distributed processing chains several Chipper instances for workload distribution, and the reasoning demo says that for models such as DeepSeek-R1 the thinking output is hidden in the web interface while the reasoning steps stay in the console output. Set against all of that, a note states the project is personal, is not designed for commercial or production use, and asks anyone intending a production deployment to do their own due diligence. Two more entries in the same list round out the feature set: audio transcription that turns audio files into text, and a client side text to speech generator built on WebAssembly, so the spoken output of a query never needs a server round trip.
Editorial conclusion
Chipper rewards a reader who opens setup.py instead of the documentation site, because the launcher, the three Ollama URL constants and the reachability probe are where the real behaviour lives, and the file states plainly that it is a personal project not designed for commercial or production use. Before building on it, work out which of the three Ollama URL branches your host and GPU profile will take, on Linux especially, where the comment in the source admits one of them does not apply. Check the release you are pinning, because the newest tag is from June 2025 and the tag naming is not consistent. And if a passing test suite is part of your decision, note that an automated unit testing framework is still an open roadmap box rather than something in the tree.
Frequently asked questions
What is the TilmanGriesel/chipper project?
Chipper is a containerised Python service for RAG pipelines, document splitting, web scraping and audio transcription, built on Haystack, Ollama, Elasticsearch and TailwindCSS. It exposes a web interface and a CLI, and it can also proxy an Ollama instance so third party clients get server side embeddings.
How do I install Chipper?
The README does not contain an install command. Its Installation and Setup section holds two links, one to a quickstart anchor and one to the project website, and the documentation is built with VitePress. The one example directory in the repository is examples/docker_compose/, and the root holds docker/, services/, a run file and run.sh.
How does Chipper decide which Ollama URL to use?
setup.py defines three defaults: http://ollama:11434 for another container, http://host.docker.internal:11434 for the host, and http://localhost:11434 for the same machine. An external URL is treated as required when the platform is Darwin, or the GPU profile is cpu, or the profile is amd-linux, and a comment in the source says host.docker.internal is not linux compatible.
Does Chipper have automated tests?
Not yet as a framework. The README roadmap lists an automated unit testing framework under upcoming features with an empty checkbox, while completed milestones already claim enhanced linting and code formatting. The root does carry .flake8, .editorconfig and .pre-commit-config.yaml.
Which Ollama clients does Chipper support?
The README names Enchanted and Open WebUI as the clients it proxies, and the third party section says any Ollama client gains server side knowledge base embeddings, model selection, query parameters and system prompt overrides. The completed roadmap marks mirroring the Ollama chat API as done.
What is the latest Chipper release?
The newest release is v2.5.0 from 2025-06-04, before that Chipper 2.4.0 from 2025-03-29, and before that a tag written v.2.3.0 from 2025-02-09. The default branch main last moved on 2026-05-19, so the repository has been worked on well after the last cut release.
Official sources
Add this badge to your README
If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.
[](https://hysenlabs.com/projects/tilmangriesel-chipper)Star history chart for your README
[](https://hysenlabs.com/projects/tilmangriesel-chipper)Was this useful?