The Retrieval-Augmented Generation use case now holds 25 providers and 98 APIs. Five of those 25 are not providers. They are catalog topic pages named database, hypermedia, support, bots and machine-learning, each with 0 APIs and a composite of 9.1 to 9.6. Meanwhile Pinecone, the vector database most RAG tutorials start with, scores 59.2, strong, on 11 APIs and is not on the page.
When we walked this use case in August it held 20 providers. The list has grown by five, and the five it added are the noise.
What the page holds
| Band | Count | Members |
|---|---|---|
| Exemplar | 4 | OpenAI 84.4, Merge 72.3, Confluence 69.1, Airbyte 68.8 |
| Strong | 3 | AIMLAPI 63.3, Amazon Nova 60.2, Azure Databricks 57.9 |
| Developing | 7 | Wikidata, Chroma 51.5, Vespa 51.1, Weaviate 50.9, Microsoft Copilot, Marqo, Apache Solr |
| Thin | 4 | Apache Jena, IEEE, MCP, Apache Lucene |
| Emerging | 2 | Ragas, Evals |
| Minimal | 5 | database, hypermedia, support, bots, machine-learning |
The August finding still holds at the top. A model provider, a unified API, a wiki and a data pipeline outrank every retrieval engine, because the hard part of RAG is getting permissioned content out of source systems, not storing vectors. Confluence and Airbyte have both slipped since August, Confluence from 79.3 to 69.1 and Airbyte from 75.7 to 68.8, but they are still exemplar.
The middle has moved. Weaviate was 61.5 on 22 APIs in August. Its record now shows 50.9 on 1 API. I have not traced whether that is a recount or a real loss, but either way it puts Chroma, Vespa and Weaviate inside the same 0.6-point spread in the developing band.
Who is missing
The catalog holds most of the retrieval stack, and the use case does not list it:
| Provider | Composite | Band | APIs |
|---|---|---|---|
| Pinecone | 59.2 | strong | 11 |
| Vectara | 48.4 | developing | 1 |
| Milvus | 38.7 | thin | 1 |
| Qdrant | 38.1 | thin | 1 |
| Cohere | 38.0 | thin | 9 |
| Ragie | 36.1 | thin | 1 |
| LlamaIndex | 35.8 | thin | 4 |
| Unstructured | 34.8 | thin | 3 |
Pinecone would be the best-scored retrieval engine on the page, above Azure Databricks. Ragie and Vectara sell RAG as a product. Unstructured and LlamaIndex do the parsing and chunking step. Cohere and Voyage AI (23.5) sell the embeddings and rerankers. A build path drawn from this page today skips the embedding, chunking and managed-RAG layers entirely.
Takeaway
This is our defect. The use-case matcher is pulling in topic pages that have no APIs, which should be filtered out by api_count: 0 alone, and it is missing providers whose own tags say vector search and embeddings. Pinecone is tagged RAG in its own record. For a builder, the August path still works: sources first (Confluence, Merge, Airbyte), then a retrieval engine. Add Pinecone to that shortlist, and look at Unstructured before you chunk anything by hand.
See the live use case at apis.io/use-cases/rag-knowledge/.