We replaced RAG with a virtual filesystem for our AI documentation assistant

https://www.mintlify.com/blog/how-we-built-a-virtual-filesystem-for-our-assistant

by denssumesh • 4 months ago

411 142 4 months ago

The real thing I think people are rediscovering with file system based search is that there’s a type of semantic search that’s not embedding based retrieval. One that looks more like how a librarian organizes files into shelves based on the domain.

We’re rediscovering forms of in search we’ve known about for decades. And it turns out they’re more interpretable to agents.

https://softwaredoug.com/blog/2026/01/08/semantic-search-wit...

whattheheckheck• 4 months ago

Turns out the millions of people in knowledge work arent librarians and they wing shit everywhere

wielebny• 4 months ago

Someone simply assumed at some point that RAG must be based on vector search, and everyone followed.

morkalork• 4 months ago

Doesn't have to be tho, I've had great success letting an agent loose on an Apache Lucene instance. Turns out LLMs are great at building queries.

softwaredoug3• 4 months ago

It’s something of a historical accident

We started with LLMs when everyone in search was building question answering systems. Those architectures look like the vector DB + chunking we associate with RAG.

Agents ability to call tools, using any retrieval backend, call that into question.

We really shouldn’t start RAG with the assumption we need that. I’ll be speaking about the subject in a few weeks

https://maven.com/p/7105dc/rag-is-the-what-agentic-search-is...

bluegatty2• 4 months ago

It was the terminology that did that more than anything. The term 'RAG' just has a lot of consequential baggage. Unfortunately.

ivanovm• 4 months ago

I don't think this was a simple assumption. LLMs used to be much dumber! GPT-3 era LLMS were not good at grep, they were not that good at recovering from errors, and they were not good at making followup queries over multiple turns of search. Multiple breakthroughs in code generation, tool use, and reasoning had to happen on the model side to make vector-based RAG look like unnecessary complexity

darkteflon4• 4 months ago

Certainly a lot of blog posts followed. Not sure that “everyone” was so blinkered.

graemefawcett• 4 months ago

RAG is like when you want someone to know something they're not quite getting so you yell a bit louder. For a workflow that's mainly search based, it's useful to keep things grounded.

Less useful in other contexts, unless you move away from traditional chunked embeddings and into things like graphs where the relationships provide constraints as much as additional grounding

safety1st2• 4 months ago

We were given a demo of a vector based approach, and it didn't work. They said our docs were too big and for some reason their chunking process was failing. So we ended up using a good old fashioned Elastic backend because that's what we know, and simply forwarding a few of these giant documents to the LLM verbatim along with the user's question. The results have been great, not a single complaint about accuracy, results are fast and cheap using OpenAI's micro models, Elastic is mature tech everyone understands so it's easy to maintain.

I think this turned out to be one of those lessons about premature optimization. It didn't need to be as complex as what people initially assumed. Perhaps with older models it would have been a different story.

khalic2• 4 months ago

This kind of circles back to ontological NLP, that was using knowledge representation as a primitive for language processing. There is _a ton_ of work in that direction.

softwaredoug3• 4 months ago

Exactly. And LLMs supervised by domain experts unlock a lot of capabilities to help with these types of knowledge organization problems.

UltraSane• 4 months ago

Inverted indexes have the major advantages of supporting Boolean operators.

skeptrune13• 4 months ago

I think it's cool that LLMs can effectively do this kind of categorization on the fly at relatively large scale. When you give the LLM tools beyond just "search", it really is effectively cheating.

czhu12• 4 months ago

Similar effort with PageIndex [1], which basically creates a table of contents like tree. Then an LLM traverses the tree to figure out which chunks are relevant for the context in the prompt.

1: https://github.com/VectifyAI/PageIndex

woah• 4 months ago

My intuition is that since AI assistants are fictional characters in a story being autocompleted by an LLM, mechanisms that are interpretable as human interactions with language and appear in the pretraining data have a surprising advantage over mechanisms that are more like speculation about how the brain works or abstract concepts.

reactordev• 4 months ago

This is also why LLMs get 80% of the way there and crap out on logic. They were trained on all the open source abandonware on GitHub.

_boffin_• 4 months ago

And next, we’ll get to tag based file systems

rao-v• 4 months ago

I got to say people also seem to be missing really simple tricks with RAG that help. Using longer chunks and appending the file path to the chunk makes a big difference.

Having said that, generally agree that keyword searching via rg and using the folder structure is easier and better.

3abiton• 4 months ago

I got to say people also seem to be missing really simple tricks with RAG that help. Using longer chunks and appending the file path to the chunk makes a big difference. > Having said that, generally agree that keyword searching via rg and using the folder structure is easier and better.

It depends on the task no? Codebase RAG for example has arguably a different setup than text search. I wonder how much the FS "native" embedding would help.

baby• 4 months ago

Yep, I was using RAG for all sorts of stuff and now moved everything to just rg+fd+cd+ls, much faster, easier, etc.

siva7• 4 months ago

Our documentation was already indexed, chunked, and stored in a Chroma database to power our search, so we built ChromaFs

It's obvious by that sentence that these guys neither understand RAG nor realized that the solution to their agentic problem didn't need any of this further abstractions including vector or grep

risyachka• 4 months ago

more and more often you see "new discoveries" that are very old concepts. the only discovery that usually happens there is that the author discovers for himself this concept. but it is essential nowadays to post it like if you discovered something new

andai• 4 months ago

I spent a while working on a retrieval system for LLMs and ended up reinventing a concordance (which is like an index).

It's basically the same thing as Google's inverted index, which is how Google search works.

Nothing new under the sun :)

stingraycharles• 4 months ago

Aren’t most successful RAGs using a combination of embedding similarity + BM25 + reranking? I thought there were very few RAGs that only did pure embedding similarity, but I may be mistaken.

manunamz• 4 months ago

Exactly. Traditional library science truly captured deep patterns of information architecture.

https://x.com/wibomd/status/1818305066303910006

Pixar got this right in Ralph Wrecks The Internet.

https://x.com/wibomd/status/1827067434794127648

seanlinehan• 4 months ago

This is definitely the way. There are good use cases for real sandboxes (if your agent is executing arbitrary code, you better it do so in an air-gapped environment).

But the idea of spinning up a whole VM to use unix IO primitives is way overkill. Makes way more sense to let the agent spit our unix-like tool calls and then use whatever your prod stack uses to do IO.

skeptrune13• 4 months ago

100% agree. However, if there were no resource tradeoffs, then a FUSE mount would probably be the way to go.

maille3• 4 months ago

Let's say I want a free, local or free-tier-llm, simple solution to search information mostly from my emails and a little bit from text, doc and pdf files. Are there any tool I should try to have ollamma or gemini able to reply with my own knowledge base?

ghywertelling2• 4 months ago

https://onyx.app/

This could be useful.

maille3• 4 months ago

Are you using it? I will definitely give it a shot, any pointers to online resources will be appreciated

maille3• 3 months ago

If you are still ready comments, have you heard of this? https://docs.kiln.tech/docs/documents-and-search-rag

pboulos2• 4 months ago

I think this is a great approach for a startup like Mintlify. I do have skepticism around how practical this would be in some of the “messier” organisations where RAG stands to add the most value. From personal experience, getting RAG to work well in places where the structure of the organisation and the information contained therein is far from hierarchical or partition-able is a very hard task.

khalic2• 4 months ago

The use case is well defined here, let’s not jump the gun. Text search, like with code, is a relatively simple problem compared to intrinsic semantic content in a book for example. I think the moral here is that RAG is not a silver bullet, the claude code team came to the same conclusion.

pboulos2• 4 months ago

I agree with your assessment.

dominotw2• 4 months ago

he claude code team came to the same conclusion.

github copilot uses rag

skeptrune13• 4 months ago

Modern OCR tooling is quite good. If the knowledge you are adding into your search database is able to be OCR'd then I think the approach we took here is able to be generalized.

mandeepj• 4 months ago

even a minimal setup (1 vCPU, 2 GiB RAM, 5-minute session lifetime) would put us north of $70,000 a year based on Daytona's per-second sandbox pricing ($0.0504/h per vCPU, $0.0162/h per GiB RAM)

$70k?

how about if we round off one zero? Give us $7000.

That number still seems to be very high.

lstodd2• 4 months ago

Hm. I think a dedicated 16-core box with 64 ram can be had for under $1000/year.

It being dedicated there are no limits on session lifetime and it'd run 16 those sessions no problem, so the real price should be around ~$70/year for that load.

lelanthran3• 4 months ago

It looks like, to me, that someone spent a long back-and-forth with an LLM refining a design - everything they wrote screams "over-engineered, lots of moving parts, creating tiny little sub-problems that need to then be solved".

I find it very hard to believe that a human designed their process around a "Daytona Sandbox" (whatever the fuck that is) at 100x markup over simply renting a VPS (a DO droplet is what, $6/m? $5/m?) and either containerising it or using FreeBSD with jails.

I'm looking at their entire design and thinking that, if I needed to do some stuff like this, I'd either go with a FUSE-based design or (more flexible) perform interceptions using LD_PRELOAD to catch exec, spawn, open, etc.

What sort of human engineer comes up with this sort of approach?

all2• 4 months ago

At that point I would buy an old mini PC off of ebay and just put it on my desk.

dmix2• 4 months ago

This puts a lot of LLM in front of the information discovery. That would require far more sophisticated prompting and guardrails. I'd be curious to see how people architect an LLM->document approach with tool calling, rather than RAG->reranker->LLM. I'm also curious what the response times are like since it's more variable.

skeptrune13• 4 months ago

Hmmm, the post is an attempt to explain that Mintlify migrated from embedding-retrieval->reranker->LLM to an agent loop with access to call POSIX tools as it desires. Perhaps we didn't provide enough detail?

dmix2• 4 months ago

That matches what I'm curious about. Where an LLM is doing the bulk of information discovery and tool calling directly. Most simpler RAGs have an LLM on the frontend mostly just doing simpler query clean up, subqueries and taxonomy, then again later to rerank and parse the data. So I'd imagine the prompting and guardrails part is much more complicated in an agent loop approach, since it's more powerful and open ended.

Galanwe2• 4 months ago

I am not familiar with the tech stack they use, but from an outsider point of view, I was sort of expecting some kind of fuse solution. Could someone explain why they went through a fake shell? There has to be a reason.

skeptrune13• 4 months ago

100% agree a FUSE mount would be the way to go given more time and resources.

Putting Chroma behind a FUSE adapter was my initial thought when I was implementing this but it was way too slow.

I think we would also need to optimize grep even if we had a FUSE mount.

This was easier in our case, because we didn’t need a 100% POSIX compatibility for our read only docs use case because the agent used only a subset of bash commands anyway to traverse the docs. This also avoids any extra infra overhead or maintenance of EC2 nodes/sandboxes that the agent would have to use.

Galanwe2• 4 months ago

Makes sense, thanks for clarifying!

readitalready• 4 months ago

Yah my Claude Code agents run a ton of Python and bash scripts. You're probably missing out on a lot of tool use cases without full tool use through POSIX compatibility.

darkteflon4• 4 months ago

Did you guys look at Firecracker-based options such as E2B and Fly.io? We’ve had positive early results on latency, but yeah … too early to tell where we end up on cost.

tschellenbach• 4 months ago

I think generally we are going from vector based search, to agentic tool use, and hierarchy based systems like skills.

skeptrune13• 4 months ago

Vector search has moved from a "complete solution" to just one tool among many which you should likely provide to an agent.

ghywertelling2• 4 months ago

Agents doing retrieval has been around for quite a while

https://huggingface.co/docs/smolagents/en/examples/rag

Agentic RAG: A More Powerful Approach We can overcome these limitations by implementing an Agentic RAG system - essentially an agent equipped with retrieval capabilities. This approach transforms RAG from a rigid pipeline into an interactive, reasoning-driven process.

The innovation of the blogpost is in the retrieval step.

ctxc• 4 months ago

haha, sweet. One of the cooler things I've read lately

dust42• 4 months ago

If grep and ls do the trick, then sure you don't need RAG/embeddings. But you also don't need an LLM: a full text search in a database will be a lot more performant, faster and use less resources.

tylergetsay• 4 months ago

I dont understand the additional complexity of mocking bash when they could just provide grep, ls, find, etc tools to the LLM

skeptrune13• 4 months ago

I agree that would have been the way to go given more time and resources. However, setting up a FUSE mount would have taken significantly longer and required additional infrastructure.

wahnfrieden• 4 months ago

agents are trained on bash grep/ls/find, not on tool-calling grep/ls/find

kenforthewin2• 4 months ago

I don't get it - everybody in this thread is talking about the death of vector DBs and files being all you need. The article clearly states that this is a layer on top of their existing Chroma db.

dominotw2• 4 months ago

what value is chromadb adding in that setup

skeptrune13• 4 months ago

yea chromadb is not the point. multiple data storage solutions work

jimbokun3• 4 months ago

I assumed they had other use cases where vector search was required.

bluegatty2• 4 months ago

RAG should no have have been represented as a context tool but rather just vector querying ad an variation of search/query - and that's it.

We were bitten by our own nomenclature.

Just a small variation in chosen acronym ... may have wrought a different outcome.

Different ways to find context are welcome, we have a long way to go!

skeptrune13• 4 months ago

agreed!

HanClinto• 4 months ago

"The agent doesn't need a real filesystem; it just needs the illusion of one. Our documentation was already indexed, chunked, and stored in a Chroma database to power our search, so we built ChromaFs: a virtual filesystem that intercepts UNIX commands and translates them into queries against that same database. Session creation dropped from ~46 seconds to ~100 milliseconds, and since ChromaFs reuses infrastructure we already pay for, the marginal per-conversation compute cost is zero."

Not to be "that guy" [0], but (especially for users who aren't already in ChromaDB) -- how would this be different for us from using a RAM disk?

"ChromaFs is built on just-bash ... a TypeScript reimplementation of bash that supports grep, cat, ls, find, and cd. just-bash exposes a pluggable IFileSystem interface, so it handles all the parsing, piping, and flag logic while ChromaFs translates every underlying filesystem call into a Chroma query."

It sounds like the expected use-case is that agents would interact with the data via standard CLI tools (grep, cat, ls, find, etc), and there is nothing Chroma-specific in the final implementation (? Do I have that right?).

The author compares the speeds against the Chroma implementation vs. a physical HDD, but I wonder how the benchmark would compare against a Ramdisk with the same information / queries?

I'm very willing to believe that Chroma would still be faster / better for X/Y/Z reason, but I would be interested in seeing it compared, since for many people who already have their data in a hierarchical tree view, I bet there could be some massive speedups by mounting the memory directories in RAM instead of HDD.

[0] https://news.ycombinator.com/item?id=9224

skeptrune13• 4 months ago

We would also be super interested to see that comparison. I agree that there isn't a specific reason why Chroma would be required to build something like this.

jrm4• 4 months ago

Is this related to that thing where somehow the entire damn world forgot about the power of boolean (and other precise) searching?

sunir• 4 months ago

I am really enjoying this renaissance in CLI world applications. There's so much possible.

I'm working on a related challenge which is mounting a virtual filesystem with FUSE that mirrors my Mac's actual filesystem (over a subtree like ~/source), so I can constrain the agents within that filesystem, and block destructive changes outside their repo.

I have it so every repo has its own long-lived agent. They do get excited and start changing other repos, which messes up memory.

I didn't want to create a system user per repo because that's obnoxious, so I created a single claude system user, and I am using the virtual file system to manage permissions. My gmail repo's agent can for instance change the gmail repo and the google_auth repo, but it can't change the rag repo.

Edit: I'm publishing it here. It's still under development. https://github.com/sunir/bashguard

jdthedisciple• 4 months ago

But SQLite is notoriously 35% faster than the filesystem [0], so why not use that?

[0] https://news.ycombinator.com/item?id=14550060

tomComb• 4 months ago

And Turso has built a Virtual Filesystem on top of their SQLite.

AgentFS https://agentfs.ai/ https://github.com/tursodatabase/agentfs

Which sounds like a great idea, except that is uses NFS instead of FUSE (note that macFUSE now has a FSKit backend so FUSE seems like the best solution for both Mac and Linux).

bob10292• 4 months ago

SQLite + GPT5.4 works very well for me.

My biggest success is a Roslyn method that takes a .NET solution and converts it into a SQLite database with Files, Lines, Symbols, and References tables. I've found this approach to perform substantially better than a flat, file-based setup (i.e., like what Copilot provides in Visual Studio). Especially, for very large projects. 100+ megs of source is no problem. The relational model enables some really elegant [non]queries that would otherwise require bespoke reflection tooling or a lot more tokens consumed.

superjan• 4 months ago

Is this something you can share in more detail? Did you document a skill for the LLM to use? And with what tasks do you see most improvement?

deskamess• 4 months ago

Interesting... what is the use case for the AI that is querying it? Is it how to develop additional features for integration with your app or do you have some other use case? Code review/audit/debugging/etc. For AI developing against an API I would think an OpenAPI json file would do the trick.

Is the Roslyn method called as part of the build/publish?

devops000• 4 months ago

Why not a simple full text search in Postgres ?

r12902• 4 months ago

Right. Give the agent permissions to search across certain tables. Wonder why reinventing file based detect for the db side is just a current fad?

yieldcrv• 4 months ago

I love the multipronged attack on RAG

RIP RAG: lasted one year at a skillset that recruiters would list on job descriptions, collectively shut down by industry professionals

badgersnake• 4 months ago

So you did GraphRAG but your graph is a filesystem tree.

namxam• 4 months ago

And you did not teach it to access chroma directly, because there is no adapter? Or because it is so much better at using FS tooling?

But in the end, I would expect, that you could add a skill / instructions on how to use chromadb directly

To be honest, I have no idea what chromadb is or how it works. But building an overlay FS seems like quite lot of work.

zbyforgotpass• 4 months ago

I don't know - we are discussing techniques - like having information in files, or in a semantic database, or in a relational database - as if there was one way that could dominate all information access. But finding the right information is not one task - if the needed information is a summary of expenses from a period of time then the best source of it will be a relational database, if it is who is the head of the HR department in a particular company - then it could probably be easy found on the company intranet pages (which are kind of graph database). It does not really matter much if the searcher is a human or LLM - there are some differences in the speed, the one time useful context length and the fact that LLMs are amnesiac - but these are just parameters, the task for humans is immensely complicated and there is no one architecture and there will not be one for LLMs.

I also vibed a brainstorming note with my knowledge base system. The initial prompt: """when I read "We replaced RAG with a virtual filesystem for our AI documentation assistant (mintlify.com)" title on HackerNews - the discussion is about RAG, filesystems, databases, graphs - but maybe there is something more fundamental in how we structure the systems so that the LLM can find the information needed to answer a question. Maybe there is nothing new - people had elaborate systems in libraries even before computers - but maybe there is something. Semantic search sounds useful - but knowing which page to return might be nearly as difficult as answering the question itself - and what about questions that require synthesis from many pages? Then we have distillation - an table of content is a kind of distillation targeting the task of search. """ Then I added a few more comments and the llm linked the note with the other pages in my kb. I am documenting that - because there were many voices against posting LLM generated content and that a prompt will be enough. IMHO the prompt is not enough - because the thought was also grounded in the whole theory I gathered in the KB. And that is also kind of on topic here. Anyway - here is the vibed note: https://zby.github.io/commonplace/notes/charting-the-knowled...

nlawalker• 4 months ago

Relative to making docs accessible to AI via filesystem tools, I've been looking around to see what kinds of patterns SDK authors are using to get AI coding agents to use the freshest documentation, and Vercel is doing something interesting with their AI SDK that I haven't seen elsewhere (although maybe I just haven't looked hard enough).

The "ai" npm package includes a root-level docs folder containing .mdx versions of the docs from their site, specific to the version of the package. Their intended AI-assisted developer experience is that people discover and install their ai-sdk skill (via their npx skills tool, which supports discovery and install of skills from most any provider, not just Vercel). The SKILL.md instructs the agent to explicitly ignore all knowledge that may have been trained into its model, and to first use grep to look for docs in node_modules/ai/docs/ before searching the website.

https://github.com/vercel/ai/blob/main/skills/use-ai-sdk/SKI...

tensor2• 4 months ago

This is one of the most confusing claims I've seen in a long time. Grep and others over files would be the equivalent of an old fashioned keyword search where most RAG uses vector search. But everything else they claim about a file system just suggests that they don't know anything about databases.

I'm not familiar with how most out of the box RAG systems categorize data, but with a database you can index content literally in any way you want. You could do it like a filesystem with hierarchy, you could do it tags, or any other design you can dream up.

The search can be keyword, like grep, or vector, like rag, or use the ranking algorithms that traditional text search uses (tf-idf, BM25), or a combination of them. You don't have to use just the top X ranked documents, you could, just like grep, evaluate all results past whatever matching threshold you have.

Search is an extremely rich field with a ton of very good established ways of doing things. Going back to grep and a file system is going back to ... I don't know, the 60s level of search tech?

brap2• 4 months ago

I get what you’re saying, and you’re right, however I can also see where they’re coming from:

Empirically, agents (especially the coding CLIs) seem to be doing so much better with files, even if the tooling around them is less than ideal.

With other custom tools they instantly lose 50 IQ points, if they even bother using the tools in the first place.

tensor2• 4 months ago

Sorry, this still makes no sense. LLMs don't care about files. The way most codings systems work is that they simply provide the whole file to the LLM rather than a subset of it. That's just a choice in how you implemented your RAG search system and database. In this case the "record" is big, a file. No doubt that works for code, but it's nonsensical outside that.

E.g. for wikipedia the logical unit would likely be an article. For a book, maybe it's a chapter, or maybe it's a paragraph. You need to design the system around your content and feed the LLM an appropriate logically related set of data.

pjm331• 4 months ago

Yeah I’ve had a lot of success with agentic search against a database.

The way I think of it, the main characteristic of agentic search is just that the agent can execute many types of adhoc queries

It’s not about a file system

As I understood it early RAG systems were all about performing that search for the agent - that’s what makes that approach “non agentic”

But when I have a database that has both embeddings and full text and you can query against both of those things and I let the agent execute whatever types of queries it wants - that’s “agentic search” in my book

darkteflon4• 4 months ago

Absolutely, agentic search is much more robust to the specific implementation details of your search setup (data quality issues, too) than the early one-shot approaches were. Anyone watching Claude Code work can see this for themselves.

thefourthchime• 4 months ago

I didn't get into the details too much, but I kept thinking, why isn't he just having an agent discover things from various data sources? I've had much better success with that.

jimbokun3• 4 months ago

Isn’t this the approach described in the article?

dboreham• 4 months ago

Also odd in that most filesystems implement directories and file names as...a database. You can use a filesystem as a database but you're not being as clever as you thought.

kjgkjhfkjf• 4 months ago

Seems like it would be simpler to give the agent tools to issue ChromaDB (or SQL) queries directly, rather than giving the LLM unix-like tools that are converted into queries under the hood using a complicated proprietary setup.

r12902• 4 months ago

Agree with this. Is this current world because llms are just trained on the file system? But in a year from now we move to the db?

slp3r• 4 months ago

This feels like massive overengineering just to bypass naive chunking. Emulating a POSIX shell in TS on top of ChromaDB to do hierarchical search is going to destroy your TTFT. Every ls and grep the agent decides to run is a separate inference cycle. You're just trading RAG context-loss for severe multi-step latency

stuaxo2• 4 months ago

Could totally have FUSE over the chunks and then there is no shell emilation.

girvo2• 4 months ago

I'll be honest that's what I expected to read about!

shaial• 4 months ago

The title says you replaced RAG, but ChromaFs is still querying Chroma on every command — you replaced RAG's interface, not RAG itself. Which is actually the more interesting finding: the retrieval was never the bottleneck, the abstraction was. Agents don't need better search. They need `grep`.

functional_dev• 4 months ago

exactly, embeddings destroy information.. exact keywords, acronyms, etc. They get squashed into floats.

That is why grep still beats it for code.

I generated visual schematic of every stage of the pipeline - https://vectree.io/c/retrieval-augmented-generation-embeddin...

stuaxo2• 4 months ago

Oh that's funny, I just built a RAG and exposing the files inside the database as files seemed like the next logical steo.

I would have used Fuse if it got to that point as then it is an actual filesystem.

bitwize• 4 months ago

What if... each agent had its own virtual file system, and anything the agent needed to access was accessible as files in the filesystem?

Congratulations, you just reinvented Plan 9. I think we're going to end up reinventing a lot of things in computing that we discovered and then forgot about because Apple/Microsoft/Google couldn't monetize them, "because AI". And I don't know how to feel about that.

fudged71• 4 months ago

since the solution here doesnt appear to be open source, I think you can get something similar by asking your agents to take AgentFS and replacing the DB with ChromaDB

nithril• 4 months ago

The news headline is misleading, it does not reflect the actual article title, which is much closer to what they truly did.

They did not replaced RAG because they are still using chunk and embedding. What they changed is the interface.

jiusanzhou• 4 months ago

Clever use of just-bash to avoid the sandbox cold-start problem. The key insight here is that agents don't need a real filesystem — they need a familiar interface backed by whatever storage you already have. We're seeing the same pattern in coding agents: directory hierarchy turns out to be a surprisingly effective knowledge graph that LLMs navigate better than embedding-based retrieval, mostly because they've been heavily trained on shell interactions.

emson• 4 months ago

This is interesting as there is definitely a middle ground for agent memory. On the openclaw side you have a single MEMORY.md file on the other you have RAG and GraphRAG. I wonder if Agent memory should be more nuanced? When an agent learns something how should it promote or degrade these memory blocks - you don’t want a trading agent memorising a bad trading pattern, for example. Also the agent might want to recall semantically similar memories, but it might also want to retrieve block relationships or groups of blocks for different purposes. We’ve been exploring all these concepts with “elfmem” (sELF improving MEMory): https://github.com/emson/elfmem Would love your feedback!

ACCount37• 4 months ago

Traditional RAG is a poor fit for this generation of LLMs, because it doesn't fit the "agentic tool use" workflow at all.

Self-guided "grep on a filesystem" often beats RAG because it allows the LLM to run "closed loop" and iteratively refine its queries until it obtains results. Self-guided search loop is a superset of what methods like reranking try to do.

I don't think vector search and retrieval is dead, but the old-fashioned RAG is. Vector search would have to be reengineered to fit into the new agentic workflows, so that the advantages of agentic LLMs can compound with that of vector search - because in current day "grep vs RAG" matchups, the former is already winning on the agentic merits.

"Optimize grep-centric search" is a surprisingly reasonable stopgap in the meanwhile.

dangoldbj• 4 months ago

I think the interesting bit here is that filesystems give the model something it can actually operate on (ls, grep, etc), not just query.

kangraemin• 4 months ago

This is essentially tool use with a filesystem interface — the LLM decides what to read instead of a retrieval pipeline choosing for it. Clean idea, and it sidesteps the chunking problem entirely.

Curious about the latency though. RAG is one round trip: embed query, fetch chunks, generate. This approach seems like it needs multiple LLM calls to navigate the tree before it can answer. How many hops does it typically take, and did you have to do anything special to keep response times reasonable?

jimbokun3• 4 months ago

In their case it was competing with cloning an entire repo before starting a session which was taking 10s of seconds.

siliconc0w• 4 months ago

I'm working on a filesystem for agents for similar reasons - https://clawfs.dev - lmk if your team would like an invite..

benlm2• 4 months ago

We use both a virtual file system and RAG — they each excel in different areas. The trick with RAG is the quality of data: we use an LLM to chunk into semantically cohesive sections, as well as generate metadata (including fact triples and links to other related chunks in the document) for every chunk as well as the document as a whole. We use voyage contextual embeddings to then embed each chunk with the document and chunk metadata. Works incredibly well. At retrieval time the agent can follow chunk links if needed, as well as analyze the raw file in the vfs. High quality instruction based reranking helps a lot too! We are often looking over 10s of thousands of documents and it’d be very inefficient to have our agents analyze just the vfs without rag.

benlm2• 4 months ago

Our vfs is also pretty powerful too, though: it is all backed by postgres then projected into files/directories for our agents. They get basic grep etc but also optimized fts tools for bm25, jq, and preview tools that show representative slices of large documents. All on top of Pydantic AI.

verall• 4 months ago

hey i tried to check out your website but i'm getting cloudflare error page code 520

znnajdla• 4 months ago

The obvious way to do this is to just give the agent a real filesystem. Most harnesses solve this by spinning up an isolated sandbox and cloning the repo. We already use sandboxes for asynchronous background agents where latency is an afterthought, but for a frontend assistant where a user is staring at a loading spinner, the approach falls apart. Our p90 session creation time (including GitHub clone and other setup) was ~46 seconds.

Am I the only one who read this and thought this is fucking insane? Who in their right mind would even consider spinning up a virtual machine and cloning a repo on every search query? And if all you need is a real filesystem why would you emulate a filesystem on top of a database (Chroma)? If you need a filesystem just use an actual filesystem! This sounds like insane gymnastics just to fit a “serverless” workflow. 850,000 searches a month (less than 1 request per second) sounds like something a single raspberry pi or Mac Mini could handle.

petcat• 4 months ago

At 850,000 conversations a month, even a minimal setup (1 vCPU, 2 GiB RAM, 5-minute session lifetime) would put us north of $70,000 a year

Am I crazy or is 850,000/month of anything...not really that much? Where are you spending all your CPU cycles and memory usage?

ChromaFs is built on just-bash by Vercel Labs (shoutout Malte!), a TypeScript reimplementation of bash that supports grep, cat, ls, find, and cd

Oh.. I see.

vlmutolo• 4 months ago

If you give every agent an isolated container to use, you’re going to be paying for the reserved memory while the container is active, even if the agent isn’t doing anything.

chelm• 4 months ago

RAG provided me no way to read the content myself. I now integrate the knowledge into a static page that I can read and edit myself in Markdown. Similar to MkDocs. But after I edit the content or remove elements that are no longer true, I build a JSON file and tell the agent how to query this source.

python -c " import json, wire, pathlib d = json.loads((pathlib.Path(wire.__file__).parent / 'assets/search_index.json').read_text()) [print(e['title'], e['url']) for e in d if 'QUERY' in (e.get('body','') + e.get('title','')).lower()] "

python -c " import json, wire, pathlib d = json.loads((pathlib.Path(wire.__file__).parent / 'assets/search_index.json').read_text()) [print(e['body']) for e in d if e.get('url','') == 'PATH'] "

https://wire.wise-relations.com/use-cases/replace-rag/

ahstilde• 4 months ago

buy off the shelf: https://archil.com/

samixg• 4 months ago

cool, but where're the benchmarks?

We replaced RAG with a virtual filesystem for our AI documentation assistant

Related Stories