Vector Search and Semantic Caching #417
Conversation
|
Some broken tests at first (need to obtain keys for the integration tests with OpenAI/HuggingFace/Azure) |
|
FYI @Spartee, @tylerhutcherson, & @banker |
README.md
Outdated
|
|
||
| A `Vector<T>` is a representation of an object that can be transformed into a vector by a Vectorizer. | ||
|
|
||
| A `VectorizerAttribute` is the abstract class you use to decorate your Vector fields, it is responsible for defining the logic to convert your Vectors into Embeddings. In the package `Redis.OM.Vectorizers` we provide vectorizers for HuggingFace, OpenAI, and AzureOpenAI to allow you to easily integrate them into your workflows. |
There was a problem hiding this comment.
the phrase "convert your Vectors into Embeddings" is a bit misleading as those two terms are relatively interchangeable. I think we're essentially talking about the definition of the various vector field attributes like distance metric, data type, dims, etc? Some of those subsumed by the choice of vectorizer for sure
There was a problem hiding this comment.
changed the verbiage a bit - hopefully this is better?
README.md
Outdated
| [RedisIdField] | ||
| public string Id { get; set; } | ||
|
|
||
| [Indexed(DistanceMetric = DistanceMetric.COSINE)] |
There was a problem hiding this comment.
Can you also show how a few other vector field attributes like index type (HNSW vs FLAT) and related args are set here?
There was a problem hiding this comment.
Added a couple of other parameters for the index definition, and explained it a bit better in the modeling section.
|
|
||
| With Redis OM, the embeddings can be completely transparent to you, they are generated and bound to the `Vector<T>` when you query/insert your vectors. If however you needed your embedding after the insertion/Query, they are available at `Vector<T>.Embedding`, and be queried either as the raw bytes, as an array of doubles or as an array of floats (depending on your vectorizer). | ||
|
|
||
| #### Configuration |
There was a problem hiding this comment.
Add other vector field attribute level configuration details here too?
There was a problem hiding this comment.
Added details about index definition in the modeling section as that's more or less where it belongs (the configuration section is talking about configuring the vectorizers)
README.md
Outdated
| With the vector defined in our model, all we need to do is create Vectors of the generic type, and insert them with our model. Using our `RedisCollection`, you can do this by simply using `Insert`: | ||
|
|
||
| ```cs | ||
| var query = new OpenAIQuery |
There was a problem hiding this comment.
naming this query makes sense given the implied caching use case here. Maybe spell it out a bit so it's clear why we are inserting a "query" object into your vector database?
There was a problem hiding this comment.
Query is confusing in this context, is OpenAICompletionResult & completionResult better? (It's something that's not query that you might actually do with these embeddings lol).
slorello89
force-pushed
the
feature/vectors
branch
from
a79689c to
d33e935
Compare
…ibution of the model files)
…otnet into feature/vectors
Introduces Vector Search and Semantic caching to Redis OM .NET.
One breaking change - replace
string[]withobject[]in some key places (e.g.Execute/ExecuteAsync) as the byte arrays needed for VectorSearch need to be passed in raw. Should be transparent to anyone using the higher-level APIs within Redis OM, but anyone using those raw commands might need to make some adjustments. See README for details as to how to use the new API.