Visualising Embeddings

[gzguevara]

I would like to visualise the embeddings with UMAP and plot them against my queries to see which chuck have been retrieved. Unfortunately, I cannot access the data from index.storage_context.vector_store. Moreover, I am using HuggingFaceEmbedding(model_name="sentence-transformers/all-MiniLM-L6-v2"). And the when i query the index and look at the nodes, they do no contain the embedding. Any idea how I could approach this visualisation?

[dosubot[bot]]

🤖

Hey there @gzguevara! Good to see you again. Let's dive into this new challenge together.

Based on the information you've provided, it seems like you're trying to access the embeddings stored in index.storage_context.vector_store and visualize them using UMAP. Here's how you can approach this:

1. Accessing the Embeddings: The embeddings are stored in the embedding_dict dictionary of the SimpleVectorStoreData class. The key is the node_id and the value is the corresponding embedding. You can access this dictionary directly if you have an instance of the SimpleVectorStore class. Here's a simple example:

   vector_store = SimpleVectorStore.from_persist_path(persist_path)
   embeddings = vector_store._data.embedding_dict

   Please replace persist_path with the path where your vector store data is persisted.

2. Visualizing the Embeddings: Once you have the embeddings, you can use UMAP to reduce their dimensionality and then plot them. Here's a simple example using matplotlib:

   import umap
   import matplotlib.pyplot as plt
   
   reducer = umap.UMAP()
   embeddings_2d = reducer.fit_transform(list(embeddings.values()))
   
   plt.scatter(embeddings_2d[:, 0], embeddings_2d[:, 1])
   plt.show()

3. Embeddings in Query Results: Currently, the query results do not include the embeddings. They only include the node_ids of the matching nodes. If you want to include the embeddings in the query results, you would need to modify the query method of the Index class to include the embeddings from the vector_store. However, this would require changing the LlamaIndex codebase, which might not be ideal.

I hope this helps! If you have any further questions, please let me know.

Sources

• llama_index/vector_stores/simple.py
• tests/embeddings/test_huggingface.py

---

This response is meant to be useful and save you time. It is not meant to be a precise solution, but rather a starting point for your own research.

Help me be more useful! Please leave a 👍 if this is helpful and 👎 if it is irrelevant.

If you want to continue the conversation, start your reply with @dosu-bot.