I’ve been using MongoDB since early 2012. I have loved it since. We manage several self-hosted clusters with a few terabytes of data, used for search, analytics, and other use cases.
But recently, we started building PDF RAG apps (Retrieval Augmented Generation) as well. For the RAG, we use Qdrant Cloud, which is good. However, I’m interested in exploring the possibility of migrating it to a self-hosted or cloud-hosted MongoDB setup.
Does anyone have experience using MongoDB as a main vector database for storage and search?
1 Like
Hi @Oleksandr_Gamaniuk,
I want to caveat my response with the note that this is coming from the perspective of how the MongoDB product team thinks about this question, and that I definitely am curious to hear what other users think.
Generally we see success with folks who have a lot of familiarity with MongoDB and are looking to integrate search functionality into their apps, whether it’s lexical or vector, without needing to manage another system. We got a lot of positive feedback about the automatic syncing between your collection and your vector search index, and are always looking to help folks be more productive in more ways on Atlas. There are also newcomers to MongoDB who are looking for a single system of record for their different indexing needs and find that they can build AI or search functionality into their apps faster on top of Atlas.
Hopefully that’s helpful context for starting to think about your use case on Atlas. Another user asked more specific questions about why us vs a dedicated vector db on the OpenAI forums a few months ago, that might be a helpful resources as well as you explore your options. I’m also happy to answer specific questions you might have as you deploy and query against your first vector index.
Hey, just my 2 cents as someone who’s been using MongoDB as the main vector db with 4149.
The initial need was for a global vector store which we could use to access the memories accumulated by our 4149 AI agents. This meant indexing on data that spans traditional documents to transcripts to more bespoke reflections on what the AI agent has learnt. Vector search quickly stops producing helpful results once you accumulate docs spanning different topics/timeframes.
This meant we needed a hybrid-search solution that let us search across vector search and traditional search.
From everything we looked at, MongoDB offers the best hybrid search. This also had the upside of letting us consolidate our data to one store, which means easier maintenance (which is big for our small team).
Getting up and running was pretty straight forward for me, as I have been using MongoDB for a similar time as you. There are wrappers for popular frameworks like LangChain if that is your approach, but the default MongoDB drivers can get the job done as well.
Overall been happy for about a year now and don’t plan on switching.
4 Likes