Show HN: Doc Retrieval/RAG as a Service
Hey HN! We're Mael and Rohith, the team behind Memoire, a document retrieval pipeline "as-a-service" (https://astarlogic.com/blog/announcements/launching-memoire).
Document retrieval pipelines have many applications, from search and recommendation systems to enabling Retrieval-Augmented Generation (RAG) for LLMs. RAG is especially powerful for securely providing proprietary or internal knowledge to LLMs, improving context awareness and reducing hallucinations.
However, building a fully functional RAG pipeline is far from simple. Beyond just choosing a vector or search database, you need to:
1. Pull and parse data from various sources (PDFs, Office files, CRM records, chats, etc.)
2. Chunk and embed documents for vector search or keyword extraction.
3. Choose databases, algorithms, and sync solutions for seamless CRUD operations.
4. Ensure scaling, maintenance, and compatibility with your main application.
Even with a great developer, building this takes 3-6 weeks. That's just for a proof of concept—staying up to date with rapidly evolving industry standards requires ongoing investment.
We built Memoire to simplify all of this. Our pipeline takes care of the heavy lifting and here's how it works:
Send us your documents -> Our system analyzes and indexes them -> Start searching
Here is a demo: https://www.youtube.com/watch?v=eVD1GNVRygs
Our alpha is a self-hosted version, but soon we will propose a cloud API with volume-based pricing.
Why did we build this?
We realised that like login pages, AI has some already-solved problems where teams spend too much time on repetitive infrastructure rather than focusing on innovation. We want to build APIs as a service focused on AI use cases, so teams can deliver faster and focus on what truly matters: improving their customers' experience.
We're launching on HN to get your honest feedback. What challenges have you faced with RAG pipelines? What features or improvements would you like to see? What AI features are you building repetitively?
Thanks for checking us out—we're excited to hear what you think!
Hey, Mael here
I'll answer any questions you may have with Rohith.
We're on the Indian and UK time zones, so if you ask a question after 11 am PST, you will have to wait until tomorrow for a reply!
Hi, This is rohith, will be here and happy to answer your questions!
If Ive already ocr'd my pdf's (AWS Textract) and stored each page (3 million) as a row in my search db. Can I send those pages with meta data to your system ? After I ask the LLM the question can the results have sources (my pages location) footnoted?
If you have the content of the OCR as text, yes!
You can send to the API metadata as long as it's JSON. The results will give you back the original text, the highlight closest to your query and the metadata. You can then pipe some or all of this information to your LLM.
Quick disclaimer though, our code as-is today would not support 3M documents, we optimised for speed on our first launch, and everything is in RAM (wait until our next update).