Marigold examples / simple-rag
simple-rag
marigold registry add https://www.bayis.co.uk/marigold-examples/
marigold application install simple-rag
| Version | 0.1.0 |
|---|---|
| Models | 2 declared |
| Archive | simple-rag-0.1.0.tar.gz |
| sha256 | 600f9cb7240dbc4f48f32a757723687ffda2aa6617e31e409dcd17bc9f8ed721 |
Verifies the full local RAG path: Marigold serving both chat and embeddings, open-webui handling chunking, storage, and retrieval on top. Confirms retrieval actually happens and discriminates correctly between documents, rather than the model answering from its own knowledge.
Models
models.yaml in this directory:
sentence-transformers/all-minilm-l6-v2– text-embedding, matches open-webui’sRAG_EMBEDDING_MODELwith no further configuration.qwen/qwen3-8b– instruct model.
Running
marigold deployment start marigold-examples/simple-rag
Open http://localhost:3000.
Test documents
Three short documents, each describing an invented entity with details that cannot exist anywhere except in that file – this is deliberate. A real-world benchmark dataset (SQuAD, HotpotQA, BEIR) risks the model having seen the source material during training, which would make a correct answer ambiguous evidence: it wouldn’t prove retrieval happened, only that the model already knew it.
doc1-bramblecroft.txt:
Bramblecroft Systems was founded in 2019 by Elena Wraxford.
The company's internal support ticket system is called Nettleburn.
As of the last audit, Bramblecroft Systems had 342 employees
across offices in Hull and Ljubljana.
doc2-corvidwatch.txt:
Corvidwatch Ltd is a wildlife monitoring cooperative established in 2016.
Its founder, Tomasz Ferrenbeck, previously worked as a hydrologist.
Corvidwatch's flagship sensor platform is named Rookwire, and the
organisation currently operates 58 monitoring stations, concentrated
along the Danube tributary network.
doc3-quillmere.txt:
Quillmere Analytics is a data consultancy founded in 2021 by
Adaeze Obiako. The firm's proprietary forecasting model is called
Driftlantern. Quillmere reported revenue of 2.7 million euros in
its most recent financial year, with clients concentrated in the
insurance sector.
Upload all three together, so each test also confirms discrimination between candidates, not just retrieval from a single uploaded file.
Questions and expected results
| Question | Expected |
|---|---|
| What is Bramblecroft Systems’ ticket system called, and how many employees does it have? | Nettleburn, 342 – sourced to doc1-bramblecroft.txt only |
| Who founded Corvidwatch, and what’s their sensor platform called? | Tomasz Ferrenbeck, Rookwire – sourced to doc2-corvidwatch.txt only |
| What was Quillmere Analytics’ revenue, and what’s their forecasting model called? | 2.7 million euros, Driftlantern – sourced to doc3-quillmere.txt only |
| What is Bramblecroft Systems’ primary competitor? | Not stated in any document – correct answer is “I don’t know.” This is the negative control: it checks whether the model fabricates an answer when nothing retrieved supports one, rather than checking retrieval itself. |
What this confirms, and what it doesn’t
Confirms: Marigold’s embedding endpoint and open-webui’s retrieval pipeline are wired correctly, and retrieval discriminates between multiple candidate documents rather than defaulting to whichever was uploaded most recently.
Doesn’t confirm: chunking quality or retrieval behaviour at realistic document lengths or counts. Treat this as a plumbing check, not a quality benchmark.
Architecture note
Marigold has no vector store of its own. Retrieval, chunking, and vector storage here are all open-webui’s own ChromaDB instance, created automatically under the shared model cache directory. Marigold supplies embedding vectors on request and nothing else in this path.