My advice for building something like this: don't get hung up on a need for vector databases and embedding.
Full text search or even grep/rg are a lot faster and cheaper to work with - no need to maintain a vector database index - and turn out to work really well if you put them in some kind of agentic tool loop.
The big benefit of semantic search was that it could handle fuzzy searching - returning results that mention dogs if someone searches for canines, for example.
Give a good LLM a search tool and it can come up with searches like "dog OR canine" on its own - and refine those queries over multiple rounds of searches.
Plus it means you don't have to solve the chunking problem!
(warning! It will download ~50MB of data for the model weights and onnx runtime on first load, but should otherwise run smoothly even on a phone)
It runs a small embedding model in the browser and returns search results in "real time".
It has a few illustrative examples where semantic search returns the intended results. For example bm25 does not understand that "j lo" or "jlo" refer to Jennifer Lopez. Similarly embedding based methods can better deal with things like typos.
EDIT: search is performed over 1000 news articles randomly sampled from 2016 to 2024
Anthropic found embeddings + BM25 (keyword search) gave the best results. (Well, after contextual summarization, and fusion, and reranking, and shoving the whole thing into an LLM...)
But sadly they didn't say how BM25 did on its own, which is the really interesting part to me.
In my own (small scale) tests with embeddings, I found that I'd be looking right at the page that contained the literal words in my query and embeddings would fail to find it... Ctrl+F wins again!
It would also blow up the price/latency of Claude code if every chunk of every file had to be read into haiku->summarized->sent to an embedding model ->reindexed into a project index and that index stored somewhere. Since there’s a lot of context inherent in things like the file structure, storing the central context in Claude.md is a lot simpler. I don’t think them not using vector embeddings in the project space is anything other than an indication that it’s hard to manage embeddings in Claude code.
I think your last point captures it, for various reasons (RL, inherent structure of code) iterative grepping is unreasonably effective. Interestingly Cursor does use embedding vectors for codebase indexing:
Seems like sometimes Cursor has a better understanding of the vibe of my codebase than Claude code, maybe this is part of it. Or maybe it’s just really marginally important in codebase indexing. Vector dbs still have a huge benefit in less verifiable domains.
In my experience the semantic/lexical search problem is better understood as a precision/recall tradeoff. Lexical search (along with boolean operators, exact phrase matching, etc.) has very high precision at the expense of lower recall, whereas semantic search sits at a higher recall/lower precision point on the curve.
Yeah, that sounds about right to me. The most effective approach does appear to be a hybrid of embeddings and BM25, which is worth exploring if you have the capacity to do so.
For most cases though sticking with BM25 is likely to be "good enough" and a whole lot cheaper to build and run.
Depends on the app and how often you need to change your embeddings, but I run my own hybrid semantic/bm25 search on my MacBook Pro across millions of documents without too much trouble.
I recently came across a “prefer the most common synonym” problem, in Google Maps, while searching for a poolhall—even literally ‘billiards’ returned results for swimming pools and chlorine. I wonder if some more NOTs aren’t necessary…interested in learning about RAGs though I’m a little behind the curve.
Alternative advice: just test and see what works best for your use case. Totally agreed embeddings are often overkill. However, sometimes they really help. The flow is something like:
- Iterate over your docs to build eval data: hundreds of pairs of [synthetic query, correct answer]. Focus on content from the docs not general LLM knowledge.
- Kick off a few parallel evaluations of different RAG configurations to see what works best for your use case: BM25, Vector, Hybrid. You can do a second pass to tune parameters: embedding model, top k, re-ranking, etc.
Agree totally. I’m spending half my time focused that problem (mostly synthetic data gen with guidance), and the other half on how to optimize once it works.
No reason to try to avoid semantic search. Dead easy to implement, works across languages to some extent and the fuzziness is worth quite alot.
You're realistically going to need chunks of some kind anyway to feed the LLM, and once you got those it's just a few lines of code to get a basic persistant ChromaDB going.
In my app the best lexical search approaches completely broke my agent. For my rag system the llm would on average take 2.1 lexical searches to get the results it needed. Which wasn’t terrible but it meant sometimes it needed up to 5 searches to find it which blew up user latency. Now that I have a hybrid semantic search + lexical search it only requires 1.1 searches per result.
The problem is not using parallel tool calling or not returning a search array. We do this across large data sets and don’t see much of a problem. It also means you can swap algorithms on the fly. Building a BM25 index over a few thousand documents is not very expensive locally. Rg and grep are freeish. If you have information on folder contents you can let your agent decide at execution time based on information need.
Embeddings just aren’t the most interesting thing here if you’re running a frontier fm.
Search arrays help, but parallel tool calling assumes you’ve solved two hard problems: generating diverse query variations, and verifying which result is correct. Most retrieval doesn’t have clean verification. The better approach is making search good enough that you sidestep verification as much as possible (hopefully you are only requiring the model to make a judgment call within its search array). In my case (OpenStreetMap data), lexical recall is unstable, but embeddings usually get it right if you narrow the search space enough—and a missed query is a stronger signal to the model that it’s done something wrong.
Besides, if you could reliably verify results, you’ve essentially built an RL harness—which is a lot harder to do than building an effective search system and probably worth more.
Simon have you ever given a talk or written about this sort of pragmatism? A spin on how to achieve this with Datasette is an easy thing to imagine IMO.
Do you have a standard prompt you use for this? I have definitely seen agentic tools doing this for me, e.g., when searching the local file system, but I'm not sure if it native behaviour for tool-using LLMs or if it is coerced via prompts.
No I've not got a good only for this yet. I've found the modern models (or the Claude Code etc harness) know how to do this already by default - you can ask them a question and give them a search tool and they'll start running and iterating on searches by themselves.
This matches what I found building an AI app for kids. Started with embeddings because everyone said to, then ripped it out and went with simple keyword matching. The extra complexity wasn't worth it for my use case. Most of the magic comes from the LLM anyway, not the retrieval layer.
Burying the lede here - your solution for avoiding using vector search is either offloading to 1) user, expecting them to remember the right terms or 2) using LLM to craft the search query? And having it iterate multiple times? Holy mother of inefficiency, this agentic focus is making us all brain dead.
Vector DB's and embeddings are dead simple to figure out, implement, and maintain. Especially for a local RAG, which is the primary context here. If I want to find my latest tabular notes on some obscure game dealing with medical concepts, I should be able to just literally type that. It shouldn't require me remembering the medical terms, or having some local (or god forbid, remote) LLM iterate through a dozen combos.
FWIW I also think this is a matter of how well one structures their personal KB. If you follow strict metadata/structure and have coherent/logical writing, you'll have better chance of getting results with text matching. For someone optimizing for vector space search, and minimizing the need for upfront logical structuring, it will not work out well.
I built a simple emacs package based on this idea [0]. It works surprisingly well, but I dont know how far it scales. It's likely not as frugal from a token usage perspective.
Are multiple LLM queries faster than vector search? Even with the example "dog OR canine" that leads to two LLM inference calls vs one. LLM inference is also more expensive than vector search.
In general RAG != Vector Search though. If a SQL query, grep, full text search or other does the job then by all means. But for relevance-based search, vector search shines.
Yes, exactly. We have our AI feature configured to use our pre-existing TypeSense integration and it's stunningly competent at figuring out exactly what search queries to use across which collections in order to find relevant results.
if this is coupled with powerful search engines beyond elastic then we are getting somewhere.
other nonmonotonic engines that can find structural information are out there.
One thing I didn’t see here that might be hurting your performance is a lack of semantic chunking. It sounds like you’re embedding entire docs, which kind of breaks down if the docs contain multiple concepts. A better approach for recall is using some kind of chunking program to get semantic chunks (I like spacy though you have to configure it a bit). Then once you have your chunks you need to append context to how this chunk relates to the rest of your doc before you do your embedding. I have found anthropics approach to contextual retrieval to be very performant in my RAG systems (https://www.anthropic.com/engineering/contextual-retrieval) you can just use gpt oss 20b as the model for generation of context.
Unless I’ve misunderstood your post and you are doing some form of this in your pipeline you should see a dramatic improvement in performance once you implement this.
hey, author (not op) here. we do do semantic chunking! I think maybe I gave the impression that we don't because of the mention of aggregating context but I tested this with questions that would require aggregating context from 15+ documents (meaning 2x that in chunks), hence the comment in the post!
Is there a way to convert documents into a hierarchical connected graph data structure which references each other similar to how we use personal knowledge tools like Obsidian and ability to traverse this graph? Is GraphRag technique trying to do this exactly?
Not exactly what you’re looking for but Wilson Lin’s search engine creates a graph from the DOM for context. Here’s his write up: https://blog.wilsonl.in/search-engine/
I mean as long as they're not too long I suppose you could use just about any heuristic for grouping sources. Just seems like it would be hard to generate succinct context if you mess it up.
My advice - use same rigor as other software development for a RAG application. Have a test suite (of say 100 cases) which says for this question correct response is this. Use an LLM judge to score each of the outputs of the RAG system. Now iterate till you get a score of 85 or so. And every change of prompts and strategy triggers this check, and ensures that output of 85 is always maintained.
It depends on how you test it. I recently found that the way devs test it differs radically from how users actually use it. When we first built our RAG, it showed promising results (around 90% recall on large knowledge bases). However, when the first actual users tried it, it could barely answer anything (closer to 30%). It turned out we relied on exact keywords too much when testing it: we knew the test knowledge base, so we formulated our questions in a way that helped the RAG find what we expected it to find. Real users don't know the exact terminology used in the articles. We had to rethink the whole thing. Lexical search is certainly not enough. Sure, you can run an agent on top of it, but that blows up latency - users aren't happy when they have to wait more than a couple of seconds.
This is the gap that kills most AI features. Devs test with queries they already know the answer to. Users come in with vague questions using completely different words. I learned to test by asking my kids to use my app - they phrase things in ways I would never predict.
Ironically, pitting a LLM (ideally a completely different model) up against what you're testing, letting it write human "out of the ordinary" queries to use as test cases tend to work well too, if you don't have kids you can use as a free workforce :)
It solves some types of issues lexical search never will. For example if a user searches "Close account", but the article is named "Deleting Your Profile".
But lexical solves issues semantic never will. Searching an invoice DB for "Initech" with semantic search is near useless.
Pick a system that can do both, including a hybrid mode, then evaluate if the complexity is worth it for you.
Depends on how important keyword matching vs something more ambiguous is to your app. In Wanderfugl there’s a bunch of queries where semantic search can find an important chunk that lacks a high bm25 score. The good news is you can get all the benefits of bm25 and semantic with a hybrid ranking. The answer isn’t one or the other.
I did an experiment while learning about LLMs and llama.cpp consisting in trying to use create a Lua extension to use llama.cpp API to enhance LLMs with agent/RAG written in Lua with simple code to learn the basics and after more than 5 hours chatting with https://aistudio.google.com/prompts/new_chat?model=gemini-3-... (see the scrapped output of the whole session attached) I've got a lot far in terms of learning how to use an LLM to help develop/debug/learn about a topic (in this case agent/RAG with llama.cpp API using Lua).
I'm posting it here just in case it can help others to see and comment/improve it (it was using around 100K tokens at the end and started getting noticeable slow but still very helpful).
I'd like to have a local, fully offline and open-source software into which I can dump all our Emails, Slack, Gdrive contents, Code, and Wiki, and then query it with free form questions such as "with which customers did we discuss feature X?", producing references to the original sources.
What are my options?
I want to avoid building my own or customising a lot. Ideally it would also recommend which models work well and have good defaults for those.
This is why I built the Nextcloud MCP server, so that you can talk with your own data. Obviously this is Nextcloud-specific, but if you're using it already then this is possible now.
The default MCP server deployment supports simple CRUD operations on your data, but if you enable vector search the MCP server will begin embedding docs/notes/etc. Currently ollama and openai are supporting embeddings providers.
The MCP server then exposes tools you can use to search your docs based on semantic search and/or bm25 (via qdrant fusion) as well as generate responses using MCP sampling.
Importantly, rather than generating responses itself, the server relies on MCP sampling so that you can use any LLM/MCP client. This MCP sampling/RAG pattern is extremely powerful and it wouldn't surprise me if there was something open source that generalizes this across other data sources.
> we use Sentence Transformers (all-MiniLM-L6-v2) as our default (solid all-around performer for speed and retrieval, English-only).
Huh, interesting. I might be building a German-language RAG at some point in my future and I never even considered that some models might not support German at all. Does anyone have any experience here? Do many models underperform or not support non-English languages?
Check under the "Retrieval" section, either RTEB Multilingual or RTEB German (under language specific).
You may also want to filter for model sizes (under "Advanced Model Filters"). For instance if you are self-hosting and running on a CPU it may make sense to limit to something like <=100M parameters models.
> Do many models underperform or not support non-English languages?
Yes they do. However:
1. German is one of the more common languages to train on so more models will support it than say, Bahasa
2. There should still be a reasonable amount of multi-lingual models available. Particularly if you're OK with using proprietary models via API. AFAIK all the frontier embedding and reranking models (non open-source) are multi-lingual
When I started playing with this stuff in the GPT-4 days (8K context!), I wrote a script that would search for a relevant passage in a book, by shoving the whole book into GPT-4, in roughly context sized chunks.
I think it was like a dollar per search or something in those days. We've come a long way!
Anthropic, in their RAG article, actually say that if your thing fits in context, you should probably just put it there instead of using RAG.
I don't know where the optimal cutoff is though, since quality does suffer with long contexts. (Not to mention price and speed.)
The hardest part in RAQ is document parsing. If you only consider text then it should be ok, but once you start having tables, tables going multiple pages, charts, ignore TOC when available, footnotes … etc, that part becomes really hard and accuracy suffers to get the context regardless of what chunking do you use.
There are some patterns to help such as RAPTOR where you make ingestion content aware and instead of just ingesting content, you start using LLMs to question and summarise the content and save that to the vector database.
But reality is, having one size fits all for RAQ is not an easy task.
The issue is the ingestion (extracting the right data in the right format). This is mainly an issue in PDFs and sometimes when there are tables added as images in Docx too. You need a mix of text and OCR extraction to get the data correctly first before start chunking and adding embeddings
Interesting perspective on the use of full-text search over vector databases for RAG. I appreciate the insights on agentic tool loops and handling fuzzy searching.
When it comes to the evals for this kind of thing, is there a standard set of test data out there that one can work with to benchmark against? ie a collection of documents with questions that should result in particular documents or chunks being cited as the most relevant match.
I kinda do want to build a local RAG? I want some significant subset of Wikipedia (I assume most people know about these) on a dedicated machine with a RAG front-end. I would have then an offline Wikipedia "librarian" I could query.
But I'm lazy and assumed that someone has already built such a thing. I'm just not aware of this "Wikipedia-RAG-in-a-box".
In this case do you even need a RAG? Most models will have been trained on Wikipedia anyway.
Give Jan (https://www.jan.ai/) a try for instance. You'll need to do a bit of research as to what model will give you the best perf on your system but one of the quantized Llama or Qwen models will probably suit you well.
You can get local RAG with Anythingllm if you want minimal effort too fwiw. Pretty much plug and play. Used it for simple testing for an idea before getting into the weeds of langchain and agentic RAG.
I'm interested in the embeddings models suggested. I had some good results with nomic in a small embedding based tool I built. I also heard a few good things about qwen3-embedding, though the latency wasn't great for my usecase so I didn't pursue it much further.
Similarly, I used sqlite-vec, and was very happy with it. (if I were already using postgres I'd have gone with that, but this was more of a cli tool).
If the author is here, did you try any of those models? how would you compare the ones you did use?
> What that means is that when you're looking to build a fully local RAG setup, you'll need to substitute whatever SaaS providers you're using for a local option for each of those components.
Even starting with having "just" the documents and vector db locally is a huge first step and much more doable than going with a local LLM at the same time. I don't know any one or any org that has the resources to run their own LLM at scale.
Hopefully my new GPU will arrive tomorrow, then I can confirm myself, but if you look around online, there are lots of private people out there running their own models. A 16 GB GPU starts at 270€, which lets you run something like deepseek r.14, 32 GB GPUs start at 1200 € and then it goes further up, in model quality and price. (Top models require something like 60- 200 GB of GPU memory I think)
So for sure any medium sized company could afford to run their own LLMs, also at scale if they want to make the investment. The question is, how much they value their confidential data. (I would not trust any of the big AI companies). And you don't usually need cutting edge reasoning and coding abilities to process basic information.
Are you using the mix quantized versions? Also, there’s a setting that disables the memory allocation guardrails they put in place, which are irrelevant since macOS handles oom quite gracefully.
I built this for local RAG https://github.com/kbrisso/byte-vision it uses llama.cpp and Elasticsearch. On a laptop with 8 GB GPU it can handle a 30K token size and summarize a fairly large PDF.
For an open source, local (or cloud) vector DB, I would also recommend checking out Chroma (https://trychroma.com). It also supports full text search. Disclaimer: I work on Chroma cloud.
Doesn't seems necessary if you are using claude via bedrock or gpt via azure. At that point, its not different then sending PII through a serverless function.
Care to explain more? I understand the prompt might not be used for training, but how about sanitizing the PII from tracking or logging or memory bugs in these serverless functions
You can vibe code a local RAG with or without vectors in 5 minutes. Like another commenter pointed out, unless your corpus is huge, you do not need vectors, but hey using vectors is fun so why not.
For what its worth, I run a local first, small model, private RAG that uses LangGraph, Neo4J knowledge graphs, I swap the models around constantly. It mostly just gets called by agent tools now.
Are you using LangGraph tool nodes? I run very small, non-RLHF instruction models with maybe a 2% failure rate on response format matching tool definition. I would also guess you do not have your orchestration and pipe configured correctly.
My advice for building something like this: don't get hung up on a need for vector databases and embedding.
Full text search or even grep/rg are a lot faster and cheaper to work with - no need to maintain a vector database index - and turn out to work really well if you put them in some kind of agentic tool loop.
The big benefit of semantic search was that it could handle fuzzy searching - returning results that mention dogs if someone searches for canines, for example.
Give a good LLM a search tool and it can come up with searches like "dog OR canine" on its own - and refine those queries over multiple rounds of searches.
Plus it means you don't have to solve the chunking problem!
I created a small app that shows the difference between embedding-based ("semantic") and bm25 search:
http://search-sensei.s3-website-us-east-1.amazonaws.com/
(warning! It will download ~50MB of data for the model weights and onnx runtime on first load, but should otherwise run smoothly even on a phone)
It runs a small embedding model in the browser and returns search results in "real time".
It has a few illustrative examples where semantic search returns the intended results. For example bm25 does not understand that "j lo" or "jlo" refer to Jennifer Lopez. Similarly embedding based methods can better deal with things like typos.
EDIT: search is performed over 1000 news articles randomly sampled from 2016 to 2024
https://www.anthropic.com/engineering/contextual-retrieval
Anthropic found embeddings + BM25 (keyword search) gave the best results. (Well, after contextual summarization, and fusion, and reranking, and shoving the whole thing into an LLM...)
But sadly they didn't say how BM25 did on its own, which is the really interesting part to me.
In my own (small scale) tests with embeddings, I found that I'd be looking right at the page that contained the literal words in my query and embeddings would fail to find it... Ctrl+F wins again!
FWIW, the org decided against vector embeddings for Claude Code due in part to maintenance. See 41:05 here: https://youtu.be/IDSAMqip6ms
It would also blow up the price/latency of Claude code if every chunk of every file had to be read into haiku->summarized->sent to an embedding model ->reindexed into a project index and that index stored somewhere. Since there’s a lot of context inherent in things like the file structure, storing the central context in Claude.md is a lot simpler. I don’t think them not using vector embeddings in the project space is anything other than an indication that it’s hard to manage embeddings in Claude code.
Some agents integrate with code intelligence tools which do use embeddings, right? (As well as "mechanical" solutions like LSPs, I imagine.)
I think it's just a case of "this isn't something we need to solve, other companies solve it already and then our thing can integrate with that."
Or maybe it's really just marginal gains compared with iterative grepping. I don't know. (Still amazed how well that works!)
I think your last point captures it, for various reasons (RL, inherent structure of code) iterative grepping is unreasonably effective. Interestingly Cursor does use embedding vectors for codebase indexing:
https://cursor.com/docs/context/codebase-indexing
Seems like sometimes Cursor has a better understanding of the vibe of my codebase than Claude code, maybe this is part of it. Or maybe it’s just really marginally important in codebase indexing. Vector dbs still have a huge benefit in less verifiable domains.
who's "the org"?
No cross encoders?
In my experience the semantic/lexical search problem is better understood as a precision/recall tradeoff. Lexical search (along with boolean operators, exact phrase matching, etc.) has very high precision at the expense of lower recall, whereas semantic search sits at a higher recall/lower precision point on the curve.
Yeah, that sounds about right to me. The most effective approach does appear to be a hybrid of embeddings and BM25, which is worth exploring if you have the capacity to do so.
For most cases though sticking with BM25 is likely to be "good enough" and a whole lot cheaper to build and run.
Depends on the app and how often you need to change your embeddings, but I run my own hybrid semantic/bm25 search on my MacBook Pro across millions of documents without too much trouble.
I recently came across a “prefer the most common synonym” problem, in Google Maps, while searching for a poolhall—even literally ‘billiards’ returned results for swimming pools and chlorine. I wonder if some more NOTs aren’t necessary…interested in learning about RAGs though I’m a little behind the curve.
Alternative advice: just test and see what works best for your use case. Totally agreed embeddings are often overkill. However, sometimes they really help. The flow is something like:
- Iterate over your docs to build eval data: hundreds of pairs of [synthetic query, correct answer]. Focus on content from the docs not general LLM knowledge.
- Kick off a few parallel evaluations of different RAG configurations to see what works best for your use case: BM25, Vector, Hybrid. You can do a second pass to tune parameters: embedding model, top k, re-ranking, etc.
I build a free system that does all this (synthetic data from docs, evals, test various RAG configs without coding each version). https://docs.kiln.tech/docs/evaluations/evaluate-rag-accurac...
That's excellent advice, the only downside being that collecting that eval data remains difficult and time-consuming.
But if you want to build truly great search that's the approach to take.
Agree totally. I’m spending half my time focused that problem (mostly synthetic data gen with guidance), and the other half on how to optimize once it works.
At this point you could also optimize your agentic flow directly in DSPy using a colbert model / Ratatouille for retrieval.
Not there yet. The biggest vectors for optimizing aren’t in the agents yet (RAG method, embedding model, etc)
No reason to try to avoid semantic search. Dead easy to implement, works across languages to some extent and the fuzziness is worth quite alot.
You're realistically going to need chunks of some kind anyway to feed the LLM, and once you got those it's just a few lines of code to get a basic persistant ChromaDB going.
In my app the best lexical search approaches completely broke my agent. For my rag system the llm would on average take 2.1 lexical searches to get the results it needed. Which wasn’t terrible but it meant sometimes it needed up to 5 searches to find it which blew up user latency. Now that I have a hybrid semantic search + lexical search it only requires 1.1 searches per result.
The problem is not using parallel tool calling or not returning a search array. We do this across large data sets and don’t see much of a problem. It also means you can swap algorithms on the fly. Building a BM25 index over a few thousand documents is not very expensive locally. Rg and grep are freeish. If you have information on folder contents you can let your agent decide at execution time based on information need.
Embeddings just aren’t the most interesting thing here if you’re running a frontier fm.
Search arrays help, but parallel tool calling assumes you’ve solved two hard problems: generating diverse query variations, and verifying which result is correct. Most retrieval doesn’t have clean verification. The better approach is making search good enough that you sidestep verification as much as possible (hopefully you are only requiring the model to make a judgment call within its search array). In my case (OpenStreetMap data), lexical recall is unstable, but embeddings usually get it right if you narrow the search space enough—and a missed query is a stronger signal to the model that it’s done something wrong.
Besides, if you could reliably verify results, you’ve essentially built an RL harness—which is a lot harder to do than building an effective search system and probably worth more.
Hmm it can capture more than just single words though, e.g. meaningful phrases or paragraphs that could be written in many ways.
Simon have you ever given a talk or written about this sort of pragmatism? A spin on how to achieve this with Datasette is an easy thing to imagine IMO.
I did a livestream thing about building RAG against FTS search in Datasette last year: https://simonwillison.net/2024/Jun/21/search-based-rag/
Do you have a standard prompt you use for this? I have definitely seen agentic tools doing this for me, e.g., when searching the local file system, but I'm not sure if it native behaviour for tool-using LLMs or if it is coerced via prompts.
No I've not got a good only for this yet. I've found the modern models (or the Claude Code etc harness) know how to do this already by default - you can ask them a question and give them a search tool and they'll start running and iterating on searches by themselves.
This matches what I found building an AI app for kids. Started with embeddings because everyone said to, then ripped it out and went with simple keyword matching. The extra complexity wasn't worth it for my use case. Most of the magic comes from the LLM anyway, not the retrieval layer.
So kinda GAR - Generation-Augmented Retrieval :-)
Burying the lede here - your solution for avoiding using vector search is either offloading to 1) user, expecting them to remember the right terms or 2) using LLM to craft the search query? And having it iterate multiple times? Holy mother of inefficiency, this agentic focus is making us all brain dead.
Vector DB's and embeddings are dead simple to figure out, implement, and maintain. Especially for a local RAG, which is the primary context here. If I want to find my latest tabular notes on some obscure game dealing with medical concepts, I should be able to just literally type that. It shouldn't require me remembering the medical terms, or having some local (or god forbid, remote) LLM iterate through a dozen combos.
FWIW I also think this is a matter of how well one structures their personal KB. If you follow strict metadata/structure and have coherent/logical writing, you'll have better chance of getting results with text matching. For someone optimizing for vector space search, and minimizing the need for upfront logical structuring, it will not work out well.
My opinion on this really isn't very extreme.
Claude Code is widely regarded to be the best coding agent tool right now and it uses search, not embeddings.
I use it to answer questions about files on my computer all the time.
I built a simple emacs package based on this idea [0]. It works surprisingly well, but I dont know how far it scales. It's likely not as frugal from a token usage perspective.
0: https://github.com/dmitrym0/dm-gptel-simple-org-memory
Are multiple LLM queries faster than vector search? Even with the example "dog OR canine" that leads to two LLM inference calls vs one. LLM inference is also more expensive than vector search.
In general RAG != Vector Search though. If a SQL query, grep, full text search or other does the job then by all means. But for relevance-based search, vector search shines.
Perhaps SQLite with FTS5? Or even better, getting DuckDB into the party as it's ecosystem seems ripe for this type of work.
Yes, exactly. We have our AI feature configured to use our pre-existing TypeSense integration and it's stunningly competent at figuring out exactly what search queries to use across which collections in order to find relevant results.
if this is coupled with powerful search engines beyond elastic then we are getting somewhere. other nonmonotonic engines that can find structural information are out there.
One thing I didn’t see here that might be hurting your performance is a lack of semantic chunking. It sounds like you’re embedding entire docs, which kind of breaks down if the docs contain multiple concepts. A better approach for recall is using some kind of chunking program to get semantic chunks (I like spacy though you have to configure it a bit). Then once you have your chunks you need to append context to how this chunk relates to the rest of your doc before you do your embedding. I have found anthropics approach to contextual retrieval to be very performant in my RAG systems (https://www.anthropic.com/engineering/contextual-retrieval) you can just use gpt oss 20b as the model for generation of context.
Unless I’ve misunderstood your post and you are doing some form of this in your pipeline you should see a dramatic improvement in performance once you implement this.
hey, author (not op) here. we do do semantic chunking! I think maybe I gave the impression that we don't because of the mention of aggregating context but I tested this with questions that would require aggregating context from 15+ documents (meaning 2x that in chunks), hence the comment in the post!
Is there a way to convert documents into a hierarchical connected graph data structure which references each other similar to how we use personal knowledge tools like Obsidian and ability to traverse this graph? Is GraphRag technique trying to do this exactly?
Not exactly what you’re looking for but Wilson Lin’s search engine creates a graph from the DOM for context. Here’s his write up: https://blog.wilsonl.in/search-engine/
Ah so you’re generating context from multiple docs for your chunks? How do you decide which docs get aggregated?
Haven’t seen an answer better than “vibes” here. Especially with data across multiple domains.
I mean as long as they're not too long I suppose you could use just about any heuristic for grouping sources. Just seems like it would be hard to generate succinct context if you mess it up.
My advice - use same rigor as other software development for a RAG application. Have a test suite (of say 100 cases) which says for this question correct response is this. Use an LLM judge to score each of the outputs of the RAG system. Now iterate till you get a score of 85 or so. And every change of prompts and strategy triggers this check, and ensures that output of 85 is always maintained.
Why is it implicit that semantic search will outperform lexical search?
Back in 2023 when I compared semantic search to lexical search (tantivy; BM25), I found the search results to be marginally different.
Even if semantic search has slightly more recall, does the problem of context warrant this multi-component, homebrew search engine approach?
By what important measure does it outperform a lexical search engine? Is the engineering time worth it?
It depends on how you test it. I recently found that the way devs test it differs radically from how users actually use it. When we first built our RAG, it showed promising results (around 90% recall on large knowledge bases). However, when the first actual users tried it, it could barely answer anything (closer to 30%). It turned out we relied on exact keywords too much when testing it: we knew the test knowledge base, so we formulated our questions in a way that helped the RAG find what we expected it to find. Real users don't know the exact terminology used in the articles. We had to rethink the whole thing. Lexical search is certainly not enough. Sure, you can run an agent on top of it, but that blows up latency - users aren't happy when they have to wait more than a couple of seconds.
This is the gap that kills most AI features. Devs test with queries they already know the answer to. Users come in with vague questions using completely different words. I learned to test by asking my kids to use my app - they phrase things in ways I would never predict.
Ironically, pitting a LLM (ideally a completely different model) up against what you're testing, letting it write human "out of the ordinary" queries to use as test cases tend to work well too, if you don't have kids you can use as a free workforce :)
I build a system to do exactly this: https://docs.kiln.tech/docs/evaluations/evaluate-rag-accurac...
Basically it:
- iterates over your docs to find knowledge specific to the content
- generates hundreds of pairs of [synthetic query, correct answer]
- evaluates different RAG configurations for recall
How did you end up changing it? Creating new evals to measure the actual user experience seems easy enough, how did that inform your stack?
Totally depends on use case.
It solves some types of issues lexical search never will. For example if a user searches "Close account", but the article is named "Deleting Your Profile".
But lexical solves issues semantic never will. Searching an invoice DB for "Initech" with semantic search is near useless.
Pick a system that can do both, including a hybrid mode, then evaluate if the complexity is worth it for you.
Depends on how important keyword matching vs something more ambiguous is to your app. In Wanderfugl there’s a bunch of queries where semantic search can find an important chunk that lacks a high bm25 score. The good news is you can get all the benefits of bm25 and semantic with a hybrid ranking. The answer isn’t one or the other.
The benefit I see is you can have queries like "conversations between two scientists".
Its very dependent on use case imo
I did an experiment while learning about LLMs and llama.cpp consisting in trying to use create a Lua extension to use llama.cpp API to enhance LLMs with agent/RAG written in Lua with simple code to learn the basics and after more than 5 hours chatting with https://aistudio.google.com/prompts/new_chat?model=gemini-3-... (see the scrapped output of the whole session attached) I've got a lot far in terms of learning how to use an LLM to help develop/debug/learn about a topic (in this case agent/RAG with llama.cpp API using Lua).
I'm posting it here just in case it can help others to see and comment/improve it (it was using around 100K tokens at the end and started getting noticeable slow but still very helpful).
You can see the scrapped text for the whole seession here https://github.com/ggml-org/llama.cpp/discussions/17600
What we use: - https://github.com/ggozad/haiku.rag
Why?
- developer oriented (easy to read Python and uses pydantic-ai)
- benchmarks available
- docling with advanced citations (on branch)
- supports deep research agent
- real open source by long term committed developer not fly by night
I'd like to have a local, fully offline and open-source software into which I can dump all our Emails, Slack, Gdrive contents, Code, and Wiki, and then query it with free form questions such as "with which customers did we discuss feature X?", producing references to the original sources.
What are my options?
I want to avoid building my own or customising a lot. Ideally it would also recommend which models work well and have good defaults for those.
This is why I built the Nextcloud MCP server, so that you can talk with your own data. Obviously this is Nextcloud-specific, but if you're using it already then this is possible now.
https://github.com/cbcoutinho/nextcloud-mcp-server
The default MCP server deployment supports simple CRUD operations on your data, but if you enable vector search the MCP server will begin embedding docs/notes/etc. Currently ollama and openai are supporting embeddings providers.
The MCP server then exposes tools you can use to search your docs based on semantic search and/or bm25 (via qdrant fusion) as well as generate responses using MCP sampling.
Importantly, rather than generating responses itself, the server relies on MCP sampling so that you can use any LLM/MCP client. This MCP sampling/RAG pattern is extremely powerful and it wouldn't surprise me if there was something open source that generalizes this across other data sources.
Would love to see someone build an example using the offline wikipedia text.
Given the full text of Wikipedia is undoubtedly part of the training data, what would having it in a RAG add?
High precision recall.
It may also be cheaper to update the source (Wikipedia) with new information than to update the model.
> we use Sentence Transformers (all-MiniLM-L6-v2) as our default (solid all-around performer for speed and retrieval, English-only).
Huh, interesting. I might be building a German-language RAG at some point in my future and I never even considered that some models might not support German at all. Does anyone have any experience here? Do many models underperform or not support non-English languages?
You can refer to https://huggingface.co/spaces/mteb/leaderboard and use that to guide your selection.
Check under the "Retrieval" section, either RTEB Multilingual or RTEB German (under language specific).
You may also want to filter for model sizes (under "Advanced Model Filters"). For instance if you are self-hosting and running on a CPU it may make sense to limit to something like <=100M parameters models.
> Do many models underperform or not support non-English languages?
Yes they do. However:
1. German is one of the more common languages to train on so more models will support it than say, Bahasa
2. There should still be a reasonable amount of multi-lingual models available. Particularly if you're OK with using proprietary models via API. AFAIK all the frontier embedding and reranking models (non open-source) are multi-lingual
Yes I can confirm that,we had resorted to a multilingual embedding model back in the day. https://link.springer.com/chapter/10.1007/978-3-031-77918-3_...
When I started playing with this stuff in the GPT-4 days (8K context!), I wrote a script that would search for a relevant passage in a book, by shoving the whole book into GPT-4, in roughly context sized chunks.
I think it was like a dollar per search or something in those days. We've come a long way!
Anthropic, in their RAG article, actually say that if your thing fits in context, you should probably just put it there instead of using RAG.
I don't know where the optimal cutoff is though, since quality does suffer with long contexts. (Not to mention price and speed.)
https://www.anthropic.com/engineering/contextual-retrieval
The context size and pricing has come so far! Now the whole book fits in context, and it's like 1 cent to put the whole thing in context.
(Well, a little more with Anthropic's models ;)
The hardest part in RAQ is document parsing. If you only consider text then it should be ok, but once you start having tables, tables going multiple pages, charts, ignore TOC when available, footnotes … etc, that part becomes really hard and accuracy suffers to get the context regardless of what chunking do you use.
There are some patterns to help such as RAPTOR where you make ingestion content aware and instead of just ingesting content, you start using LLMs to question and summarise the content and save that to the vector database.
But reality is, having one size fits all for RAQ is not an easy task.
Super noob in vector embeddings: I never considered that tables would be a complexifier. (beyond defining in a parseable format for ingestion).
Do vector databases do better with long grouped text vs table formats?
The issue is the ingestion (extracting the right data in the right format). This is mainly an issue in PDFs and sometimes when there are tables added as images in Docx too. You need a mix of text and OCR extraction to get the data correctly first before start chunking and adding embeddings
Interesting perspective on the use of full-text search over vector databases for RAG. I appreciate the insights on agentic tool loops and handling fuzzy searching.
When it comes to the evals for this kind of thing, is there a standard set of test data out there that one can work with to benchmark against? ie a collection of documents with questions that should result in particular documents or chunks being cited as the most relevant match.
Yes check out haiku-rag benchmarks and evaluations
I kinda do want to build a local RAG? I want some significant subset of Wikipedia (I assume most people know about these) on a dedicated machine with a RAG front-end. I would have then an offline Wikipedia "librarian" I could query.
But I'm lazy and assumed that someone has already built such a thing. I'm just not aware of this "Wikipedia-RAG-in-a-box".
In this case do you even need a RAG? Most models will have been trained on Wikipedia anyway.
Give Jan (https://www.jan.ai/) a try for instance. You'll need to do a bit of research as to what model will give you the best perf on your system but one of the quantized Llama or Qwen models will probably suit you well.
Thank you.
Standard Ollama probably covers a lot of this
You can get local RAG with Anythingllm if you want minimal effort too fwiw. Pretty much plug and play. Used it for simple testing for an idea before getting into the weeds of langchain and agentic RAG.
I'm interested in the embeddings models suggested. I had some good results with nomic in a small embedding based tool I built. I also heard a few good things about qwen3-embedding, though the latency wasn't great for my usecase so I didn't pursue it much further.
Similarly, I used sqlite-vec, and was very happy with it. (if I were already using postgres I'd have gone with that, but this was more of a cli tool).
If the author is here, did you try any of those models? how would you compare the ones you did use?
> What that means is that when you're looking to build a fully local RAG setup, you'll need to substitute whatever SaaS providers you're using for a local option for each of those components.
Even starting with having "just" the documents and vector db locally is a huge first step and much more doable than going with a local LLM at the same time. I don't know any one or any org that has the resources to run their own LLM at scale.
It’s also just extremely viable to just host your own vector db. You just need a server with enough ram for your hnsw index.
Aren't there a bunch of models that run OK on consumer hardware now?
Hopefully my new GPU will arrive tomorrow, then I can confirm myself, but if you look around online, there are lots of private people out there running their own models. A 16 GB GPU starts at 270€, which lets you run something like deepseek r.14, 32 GB GPUs start at 1200 € and then it goes further up, in model quality and price. (Top models require something like 60- 200 GB of GPU memory I think)
So for sure any medium sized company could afford to run their own LLMs, also at scale if they want to make the investment. The question is, how much they value their confidential data. (I would not trust any of the big AI companies). And you don't usually need cutting edge reasoning and coding abilities to process basic information.
A 120gb ram MacBook Pro will run gpt-oss-120b at a very respectable clip and I’ve found it to be quite serviceable for a lot of tasks.
I bought one for this purpose, but LM Studio doesn't seem to want to run even the most quantized versions. Any suggestions?
Are you using the mix quantized versions? Also, there’s a setting that disables the memory allocation guardrails they put in place, which are irrelevant since macOS handles oom quite gracefully.
I built this for local RAG https://github.com/kbrisso/byte-vision it uses llama.cpp and Elasticsearch. On a laptop with 8 GB GPU it can handle a 30K token size and summarize a fairly large PDF.
elasticsearch is the true limitation of rag systems...
The vector search works great once you figure it out. I wanted to focus on writing the application and not have to rewrite a document store.
For an open source, local (or cloud) vector DB, I would also recommend checking out Chroma (https://trychroma.com). It also supports full text search. Disclaimer: I work on Chroma cloud.
If you end up using any of the frontier models, don't forget to protect private information in your prompts - https://github.com/deepanwadhwa/zink
Doesn't seems necessary if you are using claude via bedrock or gpt via azure. At that point, its not different then sending PII through a serverless function.
Care to explain more? I understand the prompt might not be used for training, but how about sanitizing the PII from tracking or logging or memory bugs in these serverless functions
Interesting stack. I’ve been working on doing something like this with Apple specific tech. Swiftdata is not easy to work with.
No
How does this compare with orama?
You can vibe code a local RAG with or without vectors in 5 minutes. Like another commenter pointed out, unless your corpus is huge, you do not need vectors, but hey using vectors is fun so why not.
For what its worth, I run a local first, small model, private RAG that uses LangGraph, Neo4J knowledge graphs, I swap the models around constantly. It mostly just gets called by agent tools now.
Do you run the models locally?
No local model for me manages to get function calling right.
Are you using LangGraph tool nodes? I run very small, non-RLHF instruction models with maybe a 2% failure rate on response format matching tool definition. I would also guess you do not have your orchestration and pipe configured correctly.
Glad to see all the interest in the local RAG space, it's been something I've been pushing for a while.
I just put this example together today: https://gist.github.com/davidmezzetti/d2854ed82f2d0665ec7efd...
Rust API?
Post author (not OP) here.
Would you want a Rust SDK for Skald?
Yes, none of the languages listed are very accessible from the frameworks I'm working with.
Emailing you!