Ask HN: Scaling local FAISS and LLM RAG system (356k chunks)architectural advice

[paul2495]

I’ve been building a local-only AI assistant for security analysis that uses a FAISS vector index and a local model for reasoning over parsed tool output. The current system works well, but I’m running into scaling issues as the dataset grows. Current setup: ~356k chunks FAISS (Flat index) 384-d MiniLM embeddings llama-cpp-python for inference Metadata stored in a single pickle file (~1.5GB) Tool outputs (Nmap/YARA/Volatility/etc.) parsed into structured JSON before querying

Problems I’m running into:

Metadata pickle file loads entirely into RAM

No incremental indexing — have to rebuild the FAISS index from scratch

Query performance degrades with concurrent use

Want to scale to 1M+ chunks but not sure FAISS + pickle is the right long-term architecture

My questions for those who’ve scaled local or offline RAG systems:

How do you store metadata efficiently at this scale?

Is there a practical pattern for incremental FAISS updates?

Would a vector DB (Qdrant, Weaviate, Milvus) be a better fit for offline use?

Any lessons learned from running large FAISS indexes on consumer hardware?

Not looking for product feedback — just architectural guidance from people who’ve built similar systems.

[andre-z]

FAISS is not suitable for production. The dedicated vector search solutions solve all the issues you mentioned: you just store the metadata along with vectors in JSON format. At least, with Qdrant, it works like this: https://qdrant.tech/documentation/concepts/payload/

[paul2495]

Thanksthat makes sense and it never even crossed my mind . FAISS has been great for prototyping but I'm definitely hitting the limits around metadata, updates, and operational overhead.

One thing I’m exploring now is Qdrant in embedded mode, since the tool has to run in fully air-gapped environments (no internet, no external services, distributed on a portable SSD). The embedded version runs as a simple file-based directory, similar to SQLite:

from qdrant_client import QdrantClient client = QdrantClient(path="./qdrant_data") # local-only, no server If that model works reliably, it would solve several problems FAISS creates for my use case:

incremental updates instead of full index rebuilds

storing metadata as payloads instead of a 1.5GB pickle

much easier filtering (e.g., per-source, per-customer, per-tool)

better concurrency under load

I’m still benchmarking, but curious about your experience: Have you used Qdrant’s embedded mode in a production/offline scenario? And if so, how does it behave with larger collections (500k–1M vectors) on consumer hardware?

Not dismissing FAISS — just trying to pick the right long-term architecture for an offline tool that gets updated via USB and needs to stay lightweight for the end user.