Retriever Naming

[bilgeyucel]

The receivers' names should refer to the retrieval method instead of the input type. This allows us to have specific retrievers for document stores (e.g. Chroma has ChromaQueryRetriever that refers to Chroma's query API, not Haystack's query input).

Things to pay attention
Even if there's only one retrieval method that the document store supports, let's try to state the method in the retriever's name. e.g:

• QdrantRetriever ❌
• QdrantEmbeddingRetriever ✅

ℹ️ It's quite hard to standardize the retriever naming across all databases. Let's be quite flexible here and not be afraid to introduce new retriever types.

Some future retriever examples according to this convention

• Qdrant (Qdrant Announcement about Splade)
  • QdrantSpladeRetriever -> "Splade" is not totally clear for Haystack users but aligns with Qdrant quite well
  • QdrantSparseRetriever -> clearer for Haystack users but sounds like a quite general term, might create confusion if there are multiple sparse vector methods that cannot be supported with the same Retriever
• WeaviateHybridRetriever -> This component might take text as input (not embeddings) and return retrieved documents by using Weaviate's built-in hybrid search functionality

Tasks

Preview Give feedback

1. Rename the QdrantRetriever + /retrievers folder #134
   feature request integration:qdrant +2
2. AstraRetriever -> AstraEmbeddingRetriever #394
   7 of 7
   +0
   
3. PineconeDenseRetriever -> PineconeEmbeddingRetriever [TBD] #395
   7 of 7
   +0
   
4. ChromaQueryRetriever > ChromaQueryTextRetriever #406
   7 of 7
   breaking change +1
   

[sahusiddharth]

Hi @bilgeyucel, I'm working on this

• Issue Rename the QdrantRetriever + /retrievers folder #134 I have changed the retriever name to the proposed QdrantEmbeddingRetriever and the file retriever.py is already in retriever folder, is there anything else that needs to be done?
• AstraRetriever to AstraEmbeddingRetriever : done
• About the pinecone, can you tell me the PineconeDenseRetriever -> PineconeEmbeddingRetriever decision or should I just do it for now?

[anakin87]

Some material for thinking about Sparse Embedding Retrieval (to choose the right names):

• https://docs.pinecone.io/docs/hybrid-search
• https://qdrant.tech/articles/sparse-vectors/

[bilgeyucel]

Hi @sahusiddharth, we're still working on this issue so it's on hold at the moment. Feel free to pick another issue from "Contributions Wanted" list :)

[anakin87]

We decided for this naming convention:
Document Store + Technique + "Retriever"

Dense Embeddings Retriever → EmbeddingRetriever
Sparse Embedding Retriever → SparseEmbeddingRetriever

Examples:

• ElasticsearchBM25Retriever
• ElasticsearchEmbeddingRetriever
• QdrantEmbeddingRetriever
• QdrantSparseEmbeddingRetriever

[anakin87]

Done!

Docs will be improved according to the new Naming: #6976