Faiss Reviews

My main use case for FAISS is in a retrieval-augmented generation project using it with OpenAI, where we use FAISS to store our embeddings created by the Colbert model and for retrieval as well.

In our mental generation project, FAISS fits into our workflow by providing an indexing DB that supports token-level embedding, which most classical or typically used vector DBs do not support. FAISS came into the scene because of that, and its retrieval speed is high and very scalable, making it good for our project.

The best features FAISS offers for my team include seamless integration with Colbert and the ability to use FAISS via the Ragatouille framework, which is tailor-made for using the Colbert model. Feature-wise, FAISS allows for more accurate result retrieval, and retrieval speed is also good when comparing the index size.

Regarding features, I also emphasize that the usability of FAISS is very seamless, particularly its integration with Colbert and Ragatouille.

FAISS has positively impacted my organization by helping us increase the accuracy of retrieval documents; when we store documents in token-level embedding, the accuracy will be high. Additionally, we do not need any external server to host FAISS, allowing us to integrate it with our backend framework, making it a very flexible framework.

I currently do not think there is anything to be improved based on our experience, as Faiss performs as we expected for our workflow.

I would like to see improvement in the fact that FAISS currently stores data in byte-wise files, which means we cannot see the embeddings vectors or something.

I chose 8 out of 10 because one of the drawbacks of Faiss is that it works only in-memory. If it could provide separate persistent storage without relying on in-memory, it would reduce the overhead.

I have been using Faiss for about one year.

In my experience, FAISS is a very stable product, and I have not encountered any major issues or downtime; its scalability is also good, as it can handle increased numbers of documents and workloads.

There is no customer support available for Faiss because it is an open-source product, so I have not reached out for help.

Negative

Before FAISS , I previously used MongoDB Atlas, but it didn't support token-level embedding, which is why we switched to Faiss.

I did not purchase FAISS through the AWS Marketplace because FAISS is an open-source product.

My experience with pricing, setup cost, and licensing is straightforward, as there is no cost for acquiring FAISS since it's an open-source library; the only cost we incurred is when hosting in AWS. To work with FAISS, even a single person can work with it.

Before choosing FAISS , the only other option I evaluated was PLAID, but PLAID had low community support and very few documentation resources, making FAISS a more suitable candidate due to its more extensive community support and GPU support.

My advice for others looking into using FAISS is to first consider using other more supported frameworks for vector DBs, such as Pinecone or Weaviate. If your work is not supported by those vector DBs, then FAISS is a good solution.

My company does not have any business relationship with the FAISS vendor other than being a customer, as there is no pricing or any owner since it's an open-source product.

I did not receive any gift cards or incentives for this review.

Other than FAISS, we are using other tech products such as Langgraph and Django as a framework.

I prefer not to have my company name used in the review.

I found this interview to be good, and I don't think any changes are necessary for the future.

My overall rating for FAISS is 8 out of 10.

I made some vectors for my data set. I added them and got the embeddings from Hugging Face. I made a query key-value pair for Faiss. When you want to do a query search on it, you call Faiss and send your query within it.

I used Faiss as a basic database.

I didn't know what algorithm was being learned to fetch my query. It would be beneficial if I could set a parameter and see different query mechanisms being run. I can then compare the results to see which works better for me.

I started using Faiss two weeks ago for my university project.

I rate the solution ten out of ten for stability.

I rate the solution’s scalability an eight out of ten.

Faiss is an open-source solution.

I did not face any issues integrating Faiss with other tools. I would recommend the solution to other users. Faiss has facilitated my AI-driven project very well. I recommend that other users use it for their AI projects because it provides quick query search and has a big database.

Overall, I rate the solution nine and a half out of ten.

We use Faiss for document-related tasks, particularly in the context of augmented generation per document.

We need to build many tools to streamline its integration into production environments. All the embeddings are saved in a particular location. We have to load them and start with the search in case of a query. There could be an integration with products like Spine for automated processing. Additionally, it could be more accessible for handling larger data sets.

We have used Faiss for six months in a test environment for research purposes.

The product has better performance and stability compared to Milvus.

It is easy to implement and takes an hour to complete using NPM.

It is an open-source tool.

We evaluated Milvus before. We are researching Chroma at the moment.

Faiss works quickly and efficiently with documents related to smaller data sets. I rate it a seven out of ten.