🚀 feat: Add Atlas MongoDB as an option for Vector Store

[jinzishuai]

related to https://github.com/danny-avila/LibreChat/discussions/2304

So far, only two APIs work:

• /embed
• /query

They are sufficient to make most important use case of asking a questions based on a file work (see screenshot above)

[jinzishuai]

Note on Altas Vector Search Index

I created "default" index manually with the following JSON

{
  "fields": [
    {
      "numDimensions": 1536,
      "path": "embedding",
      "similarity": "cosine",
      "type": "vector"
    },
    {
      "path": "file_id",
      "type": "filter"
    }
  ]
}

[danny-avila]

Note on Altas Vector Search Index

I created "default" index manually with the following JSON

{
  "fields": [
    {
      "numDimensions": 1536,
      "path": "embedding",
      "similarity": "cosine",
      "type": "vector"
    },
    {
      "path": "file_id",
      "type": "filter"
    }
  ]
}

does this share collection with the File collection? because maybe we can do that. this way the files and vectors can be deleted at once

[jinzishuai]

@danny-avila thanks for the comments.

• I think sharing the collection with the File collection would totally work if we are exlusively using MongoDB only. However, if the objective is to allow users flexibility to choose different vector databases while all using a MongoDB for the API layer, I think keeping them separate is probably better. In my test above, I am using a separate database with one collection.It is totally okay and easy to bring this collection to the API Mongo database.
• About your comment in the discussion on how to switch. I agree with you that some kind of environment variable flags to switch among vector databases is sufficient and there is no need to complicate things with live data migration for a switch.

[jinzishuai]

@danny-avila I think this PR is ready for a review. I've got the most basic APIs you listed working. I haven't implemented the async methods as you have done for pgvector but I think that can wait for a separate PR once this is merged.

I do have one question that I think it is best done with your decision. How do you want a user to switch between pgvector and atlas-mongo?

Right now, my .env looks like this

# # pgvector
# DB_HOST=xxx
# DB_PORT=5432
# POSTGRES_DB=librechat-pgvector
# POSTGRES_USER=postgres. 
# POSTGRES_PASSWORD= 

# atlas-mongo vector
ATLAS_MONGO_DB_URI=mongodb+srv:/
COLLECTION_NAME=testcollection

Do you want me to introduce a new variable like VECTOR_DB_TYPE=pgvector/atlas-mongo and we can default to pgvector. We'll simply ignore the other type of variables in this case?

Thanks a lot.

[danny-avila]

Thanks, it's on my list to review. Does this has to be done manually for atlas to work here?

Note on Altas Vector Search Index

I created "default" index manually with the following JSON

{
  "fields": [
    {
      "numDimensions": 1536,
      "path": "embedding",
      "similarity": "cosine",
      "type": "vector"
    },
    {
      "path": "file_id",
      "type": "filter"
    }
  ]
}

[jinzishuai]

Thanks, it's on my list to review. Does this has to be done manually for atlas to work here?

Yes, there is likely an API or CLI tool to set it up as well but the creation of vector index needs to be done.

[danny-avila]

Thanks, it's on my list to review. Does this has to be done manually for atlas to work here?

Yes, there is likely an API or CLI tool to set it up as well but the creation of vector index needs to be done.

can you do a quick write up in the read me on how to do that exactly? I'm sure I could research this but it would be nice