Most of my complex documents are, luckily, Markdown files.
I can recommend https://github.com/tobi/qmd/
. It’s a simple CLI tool for searching in these kinds of files. My previous workflow was based on fzf, but this tool gives better results and enables even more fuzzy queries. I don’t use it for code, though.
I feel local rag system , slows down my computer (I got M1 Pro 32 GB)
So I use hosted one to prevent this. My business use vector db, so created a new db to vectorize and host my knowledge base.
1. All my knowledge base is markdown files. So I split that by header tags.
2. The split is hashed and hash value is stored in SQLite
3. The hashed version is vectorized and pushed to cloud db.
4. When ever I make changes , I run a script which splits and checks hash, if it is changed the. I upsert the document. If not I don’t do anything. This helps me keep the store up to date
For search I have a cli query which searches and fetches from vector store.
For vector generation I started using Meta-LLama-3-8B in april 2024 with Python and Transformers for each text chunk on an RTX-A6000. Wow that thing was fast but noisy and also burns 500W. So a year ago I switched to an M1 Ultra and only had to replace Transformers with Apple's MLX python library. Approximately the same speed but less heat and noise. The Llama model has 4k dimensions so at fp16 thats 8 kilobyte per chunk, which I store in a BLOB column in SQLite via numpy.save(). Between running on the RTX and M1 there is a very small difference in vector output but not enough for me to change retrieval results, regenerate the vectors or change to another LLM.
For retrieval I load all the vectors from the SQlite database into a numpy.array and hand it to FAISS. Faiss-gpu was impressively fast on the RTX6000 and faiss-cpu is slower on the M1 Ultra but still fast enough for my purposes (I'm firing a few queries per day, not per minute). For 5 million chunks memory usage is around 40 GB which both fit into the A6000 and easily fits into the 128GB of the M1 Ultra. It works, I'm happy.
I've written about this (and the post was even here on HN) but mostly from the perspective of running a RAG on your infra as an organization. But I cover the general components and alternatives to Cloud services.
Don't use a vector database for code, embeddings are slow and bad for code. Code likes bm25+trigram, that gets better results while keeping search responses snappy.
The repo includes also plpgsql_bm25rrf.sql : PL/pgSQL function for hybrid search ( plpgsql_bm25 + pgvector ) with Reciprocal Rank Fusion; and Jupyter notebook examples.
I agree. Someone here posted a drop-in for grep that added the ability to do hybrid text/vector search but the constant need to re-index files was annoying and a drag. Moreover, vector search can add a ton of noise if the model isn't meant for code search and if you're not using a re-ranker.
For all intents and purposes, running gpt-oss 20B in a while loop with access to ripgrep works pretty dang well. gpt-oss is a tool calling god compared to everything else i've tried, and fast.
Used to be, but they're very complicated to operate compared to more modern alternatives and have just gotten more and more bloated over the years. Also require a bunch of different applications for different parts of the stack in order to do the same basic stuff as e.g. Meilisearch, Manticore or Typesense.
static embedding models im finding quite fast
lee101/gobed https://github.com/lee101/gobed is 1ms on gpu :) would need to be trained for code though the bigger code llm embeddings can be high quality too so its just yea about where is ideal on the pareto fronteir really , often yea though your right it tends to be bm25 or rg even for code but yea more complex solutions are kind of possible too if its really important the search is high quality
We handle ~300k customer interactions per day, so latency and precision really matter. We built an internal RAG-based portal on top of our knowledge base (basically a much better FAQ).
On the retrieval side, I built a custom search/indexing layer (Node) specifically for service traceability and discovery. It uses a hybrid approach — embeddings + full-text search + IVF-HNSW — to index and cross-reference our APIs, services, proxies and orchestration repos. The RAG pipelines sit on top of this layer, which gives us reasonable recall and predictable latency.
Compliance and observability are still a problem. Every year new vendors show up promising audits, data lineage and observability, but none of them really handle the informational sprawl of ~600 distributed systems. The entropy keeps increasing.
Lately I’ve been experimenting with a more semantic/logical KAG approach on top of knowledge graphs to map business rules scattered across those systems. The goal is to answer higher-level questions about how things actually work — Palantir-like outcomes, but with explicit logic instead of magic.
Curious if others are moving beyond “pure RAG” toward graph-based or hybrid reasoning setups.
I'm lucky enough to have 95% of my docs in small markdown markdown files so I'm just... not (+). I'm using SQLite FTS5 (full text search) to build a normal search index and using that. Well, I already had the index so I just wired it up to my mastra agents.
Each file has a short description field, so if a keyword search surfaces the doc they check the description and if it matches, load the whole doc.
This took about one hour to set up and works very well.
(+) At least, I don't think this counts as RAG. I'm honestly a bit hazy on the definition. But there's no vectordb anyway.
Well, that is what the acronym stands for. But every source I've ever seen quickly follows by noting it's retrieval backed by a vectordb. So we'd probably find an even split of people who would call this RAG or not.
Save memory vectorizes a session, summarizes it, and stores it in SQLite. Recall memory takes vector or a previous tool run id and loads the full text output. Search takes a vector array or string array and searches through the graph using fuzzy matching and vector dot products.
Well this isn’t code, but I’ve been working on a memory system for Claude Code. This portion provides semantic search over the session files in .claude/projects. It uses OpenAI for embeddings so not completely local (would be easy to modify) and storage in ChromaDB.
Claude code / codex which internally uses ripgrep, and I'm unsure if it's using parallel mode. And, project specific static analyzers.
Studies generally show when you do agentic retrieval w/ text search, that's pretty good. Adding vector retrieval and graph rag, so the typical parallel multi-retrieval followed by reranking, gives a bit of speedup and quality lift. That lines up with my local flow experience, where it is only enough that I want that for $$$$ consumer/prosumer tools, and not easy enough for DIY that I want to invest in that locally. For those who struggle with tools like spotlight running when it shouldn't, that kind of thing turns me off on the cost/benefit side.
For code, I experiment with unsound tools (semgrep, ...) vs sound flow analyzers, carefully setup for the project. Basically, ai coders love to use grep/sed for global replace refactors and other global needs, but keeps tripped up on sound flow analysis. Similar to lint and type checking, that needs to be setup for a project and taught as a skill. I'm not happy with any of my experiments here yet however :(
I am using LangChain with a SQLite database - it works pretty well on a 16G GPU, but I started running it on a crappy NUC, which also worked with lesser results.
The real lightbulb moment is when you realise the ONLY thing a RAG passes to the LLM is a short string of search results with small chunks of text. This changes it from 'magic' to 'ahh, ok - I need better search results'. With small models you cannot pass a lot of search results ( TOP_K=5 is probably the limit ), otherwise the small models 'forget context'.
It is fun trying to get decent results - and it is a rabbithole, next step I am going into is pre-summarising files and folders.
For document processing in a side project, I've been using a local all-MiniLM model with FAISS. Works well enough for semantic matching against ~50k transaction descriptions.
The real challenge wasn't model quality - it was the chunking strategy. Financial data is weirdly structured and breaking it into sensible chunks that preserve context took more iteration than expected. Eventually settled on treating each complete record as a chunk rather than doing sliding windows over raw text. The "obvious" approaches from tutorials didn't work well at all for structured tabular-ish data.
I am surprised to see very few setups leveraging LSP support. (Language Server Protocol)
It has been added to Claude Code last month.
Most setups rely on naive grep.
I've written a few terminal tools on top of Roslyn to assist Claude in code analysis for C# code. Obviously the tools are also written with the help of Claude. Worked quite well.
In my company, we build the internal chatbot based on RAG through LangChain + Milvus + LLM. Since the documents are well formatted, it is easy to do the overlapping chunking, then all those chunking data are inserted into vector db Milvus. The hybrid search (combine dense search and sparse search) is native supported in the Milvus could help us to do better retrieve. Thus the better quality answers are there.
I made a small RAG database just using Postgres. I outlined it in the blog post below. I use it for RSS Feed organisation, and searching. They are small blobs. I do the labeling using a pseudo-KNN algorithm.
Giving the LLM tools with an OData query interface has worked well for me. In C# it's pretty trivial to set up an MCP server with OData querying for an arbitrary data model. At work we have an Excel sheet with 40k rows which the LLM was able to quickly and reliably analyse using this method.
You don’t need a vector database or graph, it really depends on your existing infrastructure , file types and needs.
The newer “agent” search approach can just query a file system or api. It’s slightly slower but easier to setup and maintain as no extra infrastructure.
Data is technically a plural but nobody uses the singular and it’s being used as a singular term often - which is completely fine I think, nobody speaks Latin anyway
The Nextcloud MCP Server [0] supports Qdrant as a vectordb to store embeddings and provide semantic search across your personal documents. This enables any LLM & MCP client (e.g. claude code) into a RAG system that you can use to chat with your files.
For local deployments, Qdrant supports storing embeddings in memory as well as in a local directory (similar to sqlite) - for larger deployments Qdrant supports running as a standalone service/sidecar and can be made available over the network.
I'm using Sonnet with 1M Context Window at work, just stuffing everything in a window (it works fine for now), and I'm hoping to investigate Recursive Language Models with DSPy when I'm using local models with Ollama
For the purposes of learning, I’ve built a chatbot using ollama, streamlit, chromadb and docling. Mostly playing around with embedding and chunking on a document library.
i took a similar path, i spun up a discord bot, used ollama, pgvector, docling for random documents, and made some specialized chunking strategies for some clunkier json data. its been a little while since i messed with it, but i really did enjoy it when i was.
it all moves so fast, i wouldnt be surprised if everything i made is now crazy outdated and it was probably like 2 months ago.
To answer the question more directly, I've spent the last couple of years with a few different quant models mostly running on llama.cpp and ollama, depending. The results are way slower than the paid token api versions, but they are completely free of external influence and cost.
However the models I've tests generally turn out to be pretty dumb at the quant level I'm running to be relatively fast. And their code generation capabilities are just a mess not to be dealt with.
lee101/gobed https://github.com/lee101/gobed static embedding models so they are embedded in milliseconds and on gpu search with a cagra style on gpu index with a few things for speed like int8 quantization on the embeddings and fused embedding and search in the same kernel as the embedding really is just a trained map of embeddings per token/averaging
TL;DR:
- chunk files, index chunks
- vector/hybrid search over the index
- node app to handle requests (was the quickest to implement, LLMs understand OpenAPI well)
I can recommend https://github.com/tobi/qmd/ . It’s a simple CLI tool for searching in these kinds of files. My previous workflow was based on fzf, but this tool gives better results and enables even more fuzzy queries. I don’t use it for code, though.
So I use hosted one to prevent this. My business use vector db, so created a new db to vectorize and host my knowledge base. 1. All my knowledge base is markdown files. So I split that by header tags. 2. The split is hashed and hash value is stored in SQLite 3. The hashed version is vectorized and pushed to cloud db. 4. When ever I make changes , I run a script which splits and checks hash, if it is changed the. I upsert the document. If not I don’t do anything. This helps me keep the store up to date
For search I have a cli query which searches and fetches from vector store.
For retrieval I load all the vectors from the SQlite database into a numpy.array and hand it to FAISS. Faiss-gpu was impressively fast on the RTX6000 and faiss-cpu is slower on the M1 Ultra but still fast enough for my purposes (I'm firing a few queries per day, not per minute). For 5 million chunks memory usage is around 40 GB which both fit into the A6000 and easily fits into the 128GB of the M1 Ultra. It works, I'm happy.
Not sure how useful it is for what you need specifically: https://blog.yakkomajuri.com/blog/local-rag
Shameless plug: https://github.com/jankovicsandras/plpgsql_bm25 BM25 search implemented in PL/pgSQL ( Unlicense / Public domain )
The repo includes also plpgsql_bm25rrf.sql : PL/pgSQL function for hybrid search ( plpgsql_bm25 + pgvector ) with Reciprocal Rank Fusion; and Jupyter notebook examples.
For all intents and purposes, running gpt-oss 20B in a while loop with access to ripgrep works pretty dang well. gpt-oss is a tool calling god compared to everything else i've tried, and fast.
Can you elaborate? What makes the modern alternatives easier to operate? What makes Elasticsearch complicated?
Asking because in my experience, Elasticsearch is pretty simple to operate unless you have a huge cluster with nodes operating in different modes.
On the retrieval side, I built a custom search/indexing layer (Node) specifically for service traceability and discovery. It uses a hybrid approach — embeddings + full-text search + IVF-HNSW — to index and cross-reference our APIs, services, proxies and orchestration repos. The RAG pipelines sit on top of this layer, which gives us reasonable recall and predictable latency.
Compliance and observability are still a problem. Every year new vendors show up promising audits, data lineage and observability, but none of them really handle the informational sprawl of ~600 distributed systems. The entropy keeps increasing.
Lately I’ve been experimenting with a more semantic/logical KAG approach on top of knowledge graphs to map business rules scattered across those systems. The goal is to answer higher-level questions about how things actually work — Palantir-like outcomes, but with explicit logic instead of magic.
Curious if others are moving beyond “pure RAG” toward graph-based or hybrid reasoning setups.
This took about one hour to set up and works very well.
(+) At least, I don't think this counts as RAG. I'm honestly a bit hazy on the definition. But there's no vectordb anyway.
save_memory, recall_memory, search
Save memory vectorizes a session, summarizes it, and stores it in SQLite. Recall memory takes vector or a previous tool run id and loads the full text output. Search takes a vector array or string array and searches through the graph using fuzzy matching and vector dot products.
It’s not fancy, but it works really well. gpt-oss
https://github.com/pj4533/seance
Studies generally show when you do agentic retrieval w/ text search, that's pretty good. Adding vector retrieval and graph rag, so the typical parallel multi-retrieval followed by reranking, gives a bit of speedup and quality lift. That lines up with my local flow experience, where it is only enough that I want that for $$$$ consumer/prosumer tools, and not easy enough for DIY that I want to invest in that locally. For those who struggle with tools like spotlight running when it shouldn't, that kind of thing turns me off on the cost/benefit side.
For code, I experiment with unsound tools (semgrep, ...) vs sound flow analyzers, carefully setup for the project. Basically, ai coders love to use grep/sed for global replace refactors and other global needs, but keeps tripped up on sound flow analysis. Similar to lint and type checking, that needs to be setup for a project and taught as a skill. I'm not happy with any of my experiments here yet however :(
The real lightbulb moment is when you realise the ONLY thing a RAG passes to the LLM is a short string of search results with small chunks of text. This changes it from 'magic' to 'ahh, ok - I need better search results'. With small models you cannot pass a lot of search results ( TOP_K=5 is probably the limit ), otherwise the small models 'forget context'.
It is fun trying to get decent results - and it is a rabbithole, next step I am going into is pre-summarising files and folders.
I open sourced the code I was using - https://github.com/acutesoftware/lifepim-ai-core
The real challenge wasn't model quality - it was the chunking strategy. Financial data is weirdly structured and breaking it into sensible chunks that preserve context took more iteration than expected. Eventually settled on treating each complete record as a chunk rather than doing sliding windows over raw text. The "obvious" approaches from tutorials didn't work well at all for structured tabular-ish data.
https://github.com/anthropics/claude-code/issues/15168
https://aws.amazon.com/blogs/machine-learning/use-language-e...
The code for it is here: https://github.com/aws-samples/rss-aggregator-using-cohere-e...
The example link no longer works, as I no longer work at AWS.
The newer “agent” search approach can just query a file system or api. It’s slightly slower but easier to setup and maintain as no extra infrastructure.
https://pypi.org/project/faiss-cpu/
If the total size of your data isn't loo large...?
Data being a plural gets me.
You might have small datums but a lot of kilobytes!
For local deployments, Qdrant supports storing embeddings in memory as well as in a local directory (similar to sqlite) - for larger deployments Qdrant supports running as a standalone service/sidecar and can be made available over the network.
[0] https://github.com/cbcoutinho/nextcloud-mcp-server
https://github.com/softwaredoug/searcharray
it all moves so fast, i wouldnt be surprised if everything i made is now crazy outdated and it was probably like 2 months ago.
To answer the question more directly, I've spent the last couple of years with a few different quant models mostly running on llama.cpp and ollama, depending. The results are way slower than the paid token api versions, but they are completely free of external influence and cost.
However the models I've tests generally turn out to be pretty dumb at the quant level I'm running to be relatively fast. And their code generation capabilities are just a mess not to be dealt with.
Works well, but I didn't tested on larger scale
Question being: WHY would I be doing RAG locally?
TL;DR: - chunk files, index chunks - vector/hybrid search over the index - node app to handle requests (was the quickest to implement, LLMs understand OpenAPI well)
I wrote about it here: https://laurentcazanove.com/blog/obsidian-rag-api
Also I've got no idea what this product does, this is just a generic page of topical ai buzzwords
Don't tell me what it is, /show me why/ you built it. Then go back and keep that reasoning in, show me why I should care