Issue Summary:

When querying the HNSW index after loading it, the query results differ from those obtained when building the index from scratch. The discrepancy is observed despite using the same query and similarity functions. Debugging reveals that the entry point ID and visited nodes are consistent, but the query results are not matching expectations.

Steps to Reproduce:

1. Build the HNSW index from scratch.
2. Perform a query on the index and note the results.
3. Save the index.
4. Load the index from storage.
5. Perform the same query on the loaded index.
6. Observe that the query results differ from the initial results.

Additional Information:

• The index is saved and loaded using methods that store both node information and metadata (dimension, collection name, neighbors, entry point ID, etc.).
• The issue persists even after waiting for the loadMeta method to complete before returning from the loadIndex method.
• Debug logs have been added to the query method, showing that the entry point ID and visited nodes are consistent, but the query results are not.

Expected Behavior:

Query results after loading the index should match the results obtained when building the index from scratch, given the same query and similarity functions.

Request for Assistance:

Seeking guidance on identifying the root cause of the inconsistency in query results after loading the index and suggestions for potential debugging or corrective measures.

Give an example would be great, thanks!

First of all, thank you for creating and sharing this project with us!

The README says this project is under the MIT license and refers to a LICENSE file:

However, there's no LICENSE file in this repo.

Since you said it's MIT-licensed, please add a LICENSE file; you can use this MIT license template (the site is by GitHub): https://choosealicense.com/licenses/mit/ or you can use https://opensource.org/license/mit, but GitHub's site is providing an already properly formatted and word-wrapped text file template, which is easier to read, so I suggest using that one.