Band II · State
When it fails, it fails like “wrong answers”
Band II is layers Retrieval & Vector Stores and Fine-tuning & Training. 27 tools sit here, and the failure mode is recognisable before you have read any of them.
Start from the symptom instead
Each is an ordered checklist through the stack, cheapest check first.
Who ends up owning it
A band is a property of the stack; a role is a property of a person. They correlate but are not the same question — which is why the count is derived rather than asserted.
The layers
03Retrieval & Vector Stores
15 toolsHolds the embeddings, returns the few that matter.. Where embeddings live, and how they get retrieved.
- pgvectorVector search has become the workload rather than a side feature.
- QdrantYou want no new datastore to operate.
- WeaviateYour data is genuinely relational and joins matter more than vectors.
- ChromaYou need durability, scaling or concurrent writes.
- PineconeResidency, cost predictability or a query language you control.
- MilvusYou need a small, comprehensible system.
- TurbopufferYou need to tune the index directly.
- UnstructuredYour input is already structured.
- ElasticsearchYou need a permissive licence — the Elastic licence is not OSI-approved.
- VespaYou want a small, low-operations component.
- RerankersYou have no way to measure whether recall actually improved.
- Jina AIEmbeddings must stay in-house.
- LangChain Text SplittersStructure matters — fixed-size splitting is rarely the right boundary.
- DoclingYour documents are plain text.
- LlamaParseDocument contents cannot leave your infrastructure.
04Fine-tuning & Training
12 toolsProduces the weights that layer 1 serves.. Adapting open weights to your own data and shape.
- UnslothYou need a training stack you can audit line by line.
- AxolotlYou need to hand-tune the training loop itself.
- LLaMA-FactoryYou want a minimal, readable training script.
- PEFTYou need to change the model's capabilities, not bolt on an adapter.
- TRLYou are training from scratch rather than aligning an existing model.
- DeepSpeedThe model fits on one device — the complexity buys nothing.
- Megatron-LMYou are fine-tuning rather than pretraining.
- torchtuneYou want the breadth of a config-driven stack.
- Hugging Face TRLTraining data cannot be uploaded to a third party.
- ColabAnything reproducible — a notebook is not a training pipeline.
- ReplicateYou need custom training code or strict residency.
- Weights & Biases LaunchYou want training infrastructure you run yourself.
Indexed elsewhere, but serving this band
The problem is in band II and the tool you need is filed under a different layer. These are the crossings, each with the reason it belongs here too — a bare cross-reference reads as a mistake, so the reason is the whole payload.
Bands are a shortcut, not a taxonomy — band II covers 2 of nine layers, so the band tells you where to look without telling you how much of the stack is involved. The method says where else this index is wrong.