Retriever
Retriever
This step conducts a similarity search using your specified distance_metric on the embeddings of your chunks in the vector index.
This step does not embed the query itself and must be preceded by an embedding step.
Valid input steps: SentenceTransformerEmbedder, OpenAIEmbedder
Step Args
| Key | Value Type | Value Description |
|---|---|---|
top_k |
int |
The number of results to return1 |
metadata_json |
object |
The metadata filters you wish to apply on the embeddings in your vector DB before doing the similarity search |
distance_metric |
literal |
'cosine' or 'dotproduct' |
vector_index_name |
str |
the name of the vector index to be queried |
The metadata_json field accepts OneContext's structured query language, which is documented here:
If your retriever is followed by a "Reranker" step, in general, you should want to retrieve more results (in top_k) than you need, as the Reranker will then "Rerank" your results, and then you can retrieve your desired quantity of final results from the Reranker.
-
Results will be ordered by distance. ↩