Alternative to QMD is much needed

[olafgeibig]

The choice of QMD locks out all the OpenClaw users that run on compute and memory constrained VPS or devices like a raspberry pi. Since the project is using hosted LLMs anyway, why not also offer hosted embeddings and re-ranker?

QMD gained quite some popularity because of its ease of use, no complicated configuration to set up, but it is a very immature product. It's okay for a set of small notes but it cannot deal with larger documents. And with large, I don't mean hundreds of pages. QMD returns exactly one hit per document, even if the document contains more passages that are relevant. And that hit might be the abstract or introduction of the note.

People run quickly into OOM, especially if they don't run their claw on a Mac Mini. With all respect, I think QMD is great, but it needs to mature. And actually, there are very mature solutions that are battle proven. For instance LanceDB has all the bells and whistles, vector, BM25, reranking. You can run with local models and hosted ones so users have the choice.

[G9Pedro]

Thanks for the detailed feedback — you're raising valid points.

QMD's limitations are real: single-hit-per-document, OOM on constrained devices, and no reranking are all known pain points. We've been working on addressing this.

What's coming:

1. PR feat(plugin): v4.0.0 — in-process hybrid retrieval, cross-encoder rerank, adaptive capture #131 (open now) is a full plugin v4.0.0 rewrite that replaces QMD shell-outs with in-process TypeScript BM25 + semantic search via @huggingface/transformers. It adds cross-encoder reranking (Jina, Voyage, SiliconFlow, Pinecone), RRF fusion, recency boost, MMR diversity, and multi-scope retrieval. This directly addresses the memory/compute concerns since it doesn't need the QMD binary at all.

2. Hosted embeddings are on the roadmap — the provider system already supports OpenClaw gateway passthrough (provider=openclaw), so users can route through their existing LLM provider instead of running local models.

3. LanceDB is interesting — we'll evaluate it as a backend option. The challenge is keeping ClawVault's zero-config, markdown-native philosophy while adding more powerful retrieval.

For now: if you're on a constrained device and QMD is hitting OOM, you can disable semantic search entirely (clawvault config set qmd.enabled false) and rely on BM25 keyword search, which is fast and memory-efficient. Plugin v4.0.0 will make this much better across the board.

Keep the feedback coming — this is exactly the kind of real-world usage info that shapes priorities.