- cross-posted to:
- [email protected]
- cross-posted to:
- [email protected]
Hello!
As a handsome local AI enjoyer™ you’ve probably noticed one of the big flaws with LLMs:
It lies. Confidently. ALL THE TIME.
(Technically, it “bullshits” - https://link.springer.com/article/10.1007/s10676-024-09775-5
I’m autistic and extremely allergic to vibes-based tooling, so … I built a thing. Maybe it’s useful to you too.
The thing: llama-conductor
llama-conductor is a router that sits between your frontend (OWUI / SillyTavern / LibreChat / etc) and your backend (llama.cpp + llama-swap, or any OpenAI-compatible endpoint). Local-first (because fuck big AI), but it should talk to anything OpenAI-compatible if you point it there (note: experimental so YMMV).
I tried to make a glass-box that makes the stack behave like a deterministic system, instead of a drunk telling a story about the fish that got away.
TL;DR: “In God we trust. All others must bring data.”
Three examples:
1) KB mechanics that don’t suck (1990s engineering: markdown, JSON, checksums)
You keep “knowledge” as dumb folders on disk. Drop docs (.txt, .md, .pdf) in them. Then:
>>attach <kb>— attaches a KB folder>>summ new— generatesSUMM_*.mdfiles with SHA-256 provenance baked in- `>> moves the original to a sub-folder
Now, when you ask something like:
“yo, what did the Commodore C64 retail for in 1982?”
…it answers from the attached KBs only. If the fact isn’t there, it tells you - explicitly - instead of winging it. Eg:
The provided facts state the Commodore 64 launched at $595 and was reduced to $250, but do not specify a 1982 retail price. The Amiga’s pricing and timeline are also not detailed in the given facts.
Missing information includes the exact 1982 retail price for Commodore’s product line and which specific model(s) were sold then. The answer assumes the C64 is the intended product but cannot confirm this from the facts.
Confidence: medium | Source: Mixed
No vibes. No “well probably…”. Just: here’s what’s in your docs, here’s what’s missing, don’t GIGO yourself into stupid.
And when you’re happy with your summaries, you can:
>>move to vault— promote those SUMMs into Qdrant for the heavy mode.
2) Mentats: proof-or-refusal mode (Vault-only)
Mentats is the “deep think” pipeline against your curated sources. It’s enforced isolation:
- no chat history
- no filesystem KBs
- no Vodka
- Vault-only grounding (Qdrant)
It runs triple-pass (thinker → critic → thinker). It’s slow on purpose. You can audit it. And if the Vault has nothing relevant? It refuses and tells you to go pound sand:
FINAL_ANSWER:
The provided facts do not contain information about the Acorn computer or its 1995 sale price.
Sources: Vault
FACTS_USED: NONE
[ZARDOZ HATH SPOKEN]
Also yes, it writes a mentats_debug.log, because of course it does. Go look at it any time you want.
The flow is basically: Attach KBs → SUMM → Move to Vault → Mentats. No mystery meat. No “trust me bro, embeddings.”
3) Vodka: deterministic memory on a potato budget
Local LLMs have two classic problems: goldfish memory + context bloat that murders your VRAM.
Vodka fixes both without extra model compute. (Yes, I used the power of JSON files to hack the planet instead of buying more VRAM from NVIDIA).
!!stores facts verbatim (JSON on disk)??recalls them verbatim (TTL + touch limits so memory doesn’t become landfill)- CTC (Cut The Crap) hard-caps context (last N messages + char cap) so you don’t get VRAM spikes after 400 messages
So instead of:
“Remember my server is 203.0.113.42” → “Got it!” → [100 msgs later] → “127.0.0.1 🥰”
you get:
!! my server is 203.0.113.42?? server ip→ 203.0.113.42 (with TTL/touch metadata)
And because context stays bounded: stable KV cache, stable speed, your potato PC stops crying.
There’s more (a lot more) in the README, but I’ve already over-autism’ed this post.
TL;DR:
If you want your local LLM to shut up when it doesn’t know and show receipts when it does, come poke it:
- Primary (Codeberg): https://codeberg.org/BobbyLLM/llama-conductor
- Mirror (GitHub): https://github.com/BobbyLLM/llama-conductor
PS: Sorry about the AI slop image. I can’t draw for shit.
PPS: A human with ASD wrote this using Notepad++. If it the formatting is weird, now you know why.
You must log in or # to comment.
Holy shit I’m glad to be on the autistic side of the internet.
Thank you for proving that fucking JSON text files are all you need and not “just a couple billion more parameters bro”
Awesome work, all the kudos.
Thanks. It’s not perfect but I hope it’s a step in a useful direction
This sounds really interesting, I’m looking forward to reading the comments here in detail and looking at the project, might even end up incorporating it into my own!
I’m working on something that addresses the same problem in a different way, the problem of constraining or delineating the specifically non-deterministic behavior one wants to involve in a complex workflow. Your approach is interesting and has a lot of conceptual overlap with mine, regarding things like strictly defining compliance criteria and rejecting noncompliant outputs, and chaining discrete steps into a packaged kind of “super step” that integrates non-deterministic substeps into a somewhat more deterministic output, etc.
How involved was it to build it to comply with the OpenAI API format? I haven’t looked into that myself but may.
Cheers!
Re: OpenAI API format: 3.6 - not great, not terrible :)
In practice I only had to implement a thin subset: POST /v1/chat/completions + GET /v1/models (most UIs just need those). The payload is basically {model, messages, temperature, stream…} and you return a choices[] with an assistant message. The annoying bits are the edge cases: streaming/SSE if you want it, matching the error shapes UIs expect, and being consistent about model IDs so clients don’t scream “model not found”. Which is actually a bug I still need to squash some more for OWUI 0.7.2. It likes to have its little conniptions.
But TL;DR: more plumbing than rocket science. The real pain was sitting down with pen and paper and drawing what went where and what wasn’t allowed to do what. Because I knew I’d eventually fuck something up (I did, many times), I needed a thing that told me “no, that’s not what this is designed to do. Do not pass go. Do not collect $200”.
shrug I tried.
I’m probably going to give this a try, but I think you should make it clearer for those who aren’t going to dig through the code that it’s still LLMs all the way down and can still have issues - it’s just there are LLMs double-checking other LLMs work to try to find those issues. There are still no guarantees since it’s still all LLMs.
I haven’t tried this tool specifically, but I do on occasion ask both Gemini and ChatGPT’s search-connected models to cite sources when claiming stuff and it doesn’t seem to even slightly stop them bullshitting and claiming a source says something that it doesn’t.
Yeah, this is different. Try it. It gives you cryptogenic key to the source (which you must provide yourself: please be aware. GIGO).
How does having a key solve anything? Its not that the source doesn’t exist, it’s that the source says something different to the LLM’s interpretation of it.
Yeah.
The SHA isn’t there to make the model smarter. It’s there to make the source immutable and auditable.
Having been burnt by LLMs (far too many times), I now start from a position of “fuck you, prove it”.
The hash proves which bytes the answer was grounded in, should I ever want to check it. If the model misreads or misinterprets, you can point to the source and say “the mistake is here, not in my memory of what the source was.”.
If it does that more than twice, straight in the bin. I have zero chill any more.
Secondly, drift detection. If someone edits or swaps a file later, the hash changes. That means yesterday’s answer can’t silently pretend it came from today’s document. I doubt my kids are going to sneak in and change the historical prices of 8 bit computers (well, the big one might…she’s dead keen on being a hacker) but I wanted to be sure no one and no-thing was fucking with me.
Finally, you (or someone else) can re-run the same question against the same hashed inputs and see if the system behaves the same way.
So: the hashes don’t fix hallucinations (I don’t even think that’s possible, even with magic). The hashes make it possible to audit the answer and spot why hallucinations might have happened.
PS: You’re right that interpretation errors still exist. That’s why Mentats does the triple-pass and why the system clearly flags “missing / unsupported” instead of filling gaps. The SHA is there to make the pipeline inspectable, instead of “trust me, bro.”.
Guess what? I don’t trust you. Prove it or GTFO.
soooo if it doesn’t know something it won’t say anything and if it does know something it’ll show sources…so essentially you plug this into Claude it’s just never going to say anything to you ever again?
neat.
I see what you did there :)
Claude! Look how they massacred my boy!
I want to believe you, but that would mean you solved hallucination.
Either:
A) you’re lying
B) you’re wrong
C) KB is very small
D) None of the above.
I didn’t “solve hallucination”. I changed the failure mode. The model can still hallucinate internally. The difference is it’s not allowed to surface claims unless they’re grounded in attached sources.
If retrieval returns nothing relevant, the router forces a refusal instead of letting the model free-associate. So the guarantee isn’t “the model is always right.”
The guarantee is “the system won’t pretend it knows when the sources don’t support it.” That’s it. That’s the whole trick.
KB size doesn’t matter much here. Small or large, the constraint is the same: no source, no claim. GTFO.
That’s a control-layer property, not a model property. If it helps: think of it as moving from “LLM answers questions” to “LLM summarizes evidence I give it, or says ‘insufficient evidence.’”
Again, that’s the whole trick.
You don’t need to believe me. In fact, please don’t. Test it.
I could be wrong…but if I’m right (and if you attach this to a non-retarded LLM), then maybe, just maybe, this doesn’t suck balls as much as you think it might.
Maybe it’s even useful to you.
I dunno. Try it?
So… Rag with extra steps and rag summarization? What about facts that are not rag retrieval?
Parts of this are RAG, sure
RAG parts:
- Vault / Mentats is classic retrieval + generation.
- Vector store = Qdrant
- Embedding and reranker
So yes, that layer is RAG with extra steps.
What’s not RAG -
KB mode (filesystem SUMM path)
This isn’t vector search. It’s deterministic, file-backed grounding. You attach folders as needed. The system summarizes and hashes docs. The model can only answer from those summaries in that mode. There’s no semantic retrieval step. It can style and jazz around the answer a little, but the answer is the answer is the answer.
If the fact isn’t in the attached KB, the router forces a refusal. Put up or shut up.
Vodka (facts memory)
That’s not retrieval at all, in the LLM sense. It’s verbatim key-value recall.
- JSON on disk
- Exact store (!!)
- Exact recall (??)
Again, no embeddings, no similarity search, no model interpretation.
“Facts that aren’t RAG”
In my set up, they land in one of two buckets.
-
Short-term / user facts → Vodka. That for things like numbers, appointments, lists, one-off notes etc. Deterministic recall, no synthesis.
-
Curated knowledge → KB / Vault. Things you want grounded, auditable, and reusable.
In response to the implicit “why not just RAG then”
Classic RAG failure mode is: retrieval is fuzzy → model fills gaps → user can’t tell which part came from where.
The extra “steps” are there to separate memory from knowledge, separate retrieval from synthesis and make refusal a legal output, not a model choice.
So yeah; some of it is RAG. RAG is good. The point is this system is designed so not everything of value is forced through a semantic search + generate loop. I don’t trust LLMs. I am actively hostile to them. This is me telling my LLM to STFU and prove it, or GTFO. I know that’s a weird way to operate maybe (advesarial, assume worst, engineer around issue) but that’s how ASD brains work.
The system summarizes and hashes docs. The model can only answer from those summaries in that mode
Oh boy. So hallucination will occur here, and all further retrievals will be deterministically poisoned?
Oh boy. So hallucination will occur here, and all further retrievals will be deterministically poisoned?
Huh? That is the literal opposite of what I said. Like, diametrically opposite.
Let me try this a different way.
Hallucination in SUMM doesn’t “poison” the KB, because SUMMs are not authoritative facts, they’re derived artifacts with provenance. They’re explicitly marked as model output tied to a specific source hash. Two key mechanics that stop the cascade you’re describing:
- SUMM is not a “source of truth”
The source of truth is still the original document, not the summary. The summary is just a compressed view of it. That’s why it carries a SHA of the original file. If a SUMM looks wrong, you can:
a) trace it back to the exact document version b) regenerate it c) discard it d) read the original doc yourself and manually curate it.
Nothing is “silently accepted” as ground truth.
- Promotion is manual, not automatic
The dangerous step would be: model output -> auto-ingest into long-term knowledge.
That’s explicitly not how this works.
The Flow is: Attach KB -> SUMM -> human reviews -> Ok, move to Vault -> Mentats runs against that
Don’t like a SUMM? Don’t push it into the vault. There’s a gate between “model said a thing” and “system treats this as curated knowledge.” That’s you - the human. Don’t GI and it won’t GO.
Determinism works for you here. The hash doesn’t freeze the hallucination; it freezes the input snapshot. That makes bad summaries:
- reproducible
- inspectable
- fixable
Which is the opposite of silent drift.
If SUMM is wrong and you miss it, the system will be consistently wrong in a traceable way, not creatively wrong in a new way every time.
That’s a much easier class of bug to detect and correct. Again: the proposition is not “the model will never hallucinate.”. It’s “it can’t silently propagate hallucinations without a human explicitly allowing it to, and when it does, you trace it back to source version”.
And that, is ultimately what keeps the pipeline from becoming “poisoned”.
Huh? That is the literal opposite of what I said. Like, diametrically opposite.
The system summarizes and hashes docs. The model can only answer from those summaries in that mode. There’s no semantic retrieval step.
No, that’s exactly what you wrote.
Now, with this change
SUMM -> human reviews
That would be fixed, but will work only for small KBs, as otherwise the summary would be exhaustive.
Case in point: assume a Person model with 3-7 facts per Person. Assume small 3000 size set of Persons. How would the SUMM of work? Do you expect a human to verify that SUMM? How are you going to converse with your system to get the data from that KB Person set? Because to me that sounds like case C, only works for small KBs.
Again: the proposition is not “the model will never hallucinate.”. It’s “it can’t silently propagate hallucinations without a human explicitly allowing it to, and when it does, you trace it back to source version”.
Fair. Except that you are still left with the original problem of you don’t know WHEN the information is incorrect if you missed it at SUMM time.
Replying in specific
“SUMM -> human reviews That would be fixed, but will work only for small KBs, as otherwise the summary would be exhaustive.”
Correct: filesystem SUMM + human review is intentionally for small/curated KBs, not “review 3,000 entities.” The point of SUMM is curation, not bulk ingestion at scale. If the KB is so large that summaries become exhaustive, that dataset is in the wrong layer.
“Case in point: assume a Person model with 3-7 facts per Person. Assume small 3000 size set of Persons. How would the SUMM of work?”
Poorly. It shouldn’t work via filesystem SUMM. A “Person table” is structured data; SUMM is for documents. For 3,000 people × (3–7 facts), you’d put that in a structured store (SQLite/CSV/JSONL/whatever) and query it via a non-LLM tool (exact lookup/filter) or via Vault retrieval if you insist on LLM synthesis on top.
Do you expect a human to verify that SUMM?”
No - not for that use case. Human verification is realistic when you’re curating dozens/hundreds of docs, not thousands of structured records. For 3,000 persons, verification is done by data validation rules (schema, constraints, unit tests, diff checks), not reading summaries.
“How are you going to converse with your system to get the data from that KB Person set?”
Not by attaching a folder and “asking the model nicely.” You’d do one of these -
- Exact tool lookup: person(“Alice”) -> facts, or search by ID/name, return rows deterministically.
- Hybrid: tool lookup returns the relevant rows, then the LLM formats/summarizes them.
- Vault retrieval: embed/chunk rows and retrieve top-k, but that’s still weaker than exact lookup for structured “Person facts.”
So: conversation is fine as UX, but the retrieval step should be tool-based (exact) for that dataset.
But actually, you give me a good idea here. It wouldn’t be the work of ages to build a >>look or >>find function into this thing. Maybe I will.
My mental model for this was always “1 person, 1 box, personal scale” but maybe I need to think bigger. Then again, scope creep is a cruel bitch.
“Because to me that sounds like case C, only works for small KBs.”
For filesystem SUMM + human review: yes. That’s the design. It’s a personal, “curate your sources” workflow, not an enterprise entity store.
This was never designed to be a multi-tenant look up system. I don’t know how to build that and still keep it 1) small 2) potato friendly 3) account for ALL the moving part nightmares that brings.
What I built is STRICTLY for personal use, not enterprise use.
Fair. Except that you are still left with the original problem of you don’t know WHEN the information is incorrect if you missed it at SUMM time.”
Sort of. Summation via LLM was always going to be a lossy proposition. What this system changes is the failure mode:
- Without this: errors can get injected and later you can’t tell where they came from.
- With this: if a SUMM is wrong, it is pinned to a specific source file hash + summary hash, and you can fix it by re-summarizing or replacing the source.
In other words: it doesn’t guarantee correctness; it guarantees traceability and non-silent drift. You still need to “trust but verify”.
TL;DR:
You don’t query big, structured datasets (like 3,000 “Person” records) via SUMM at all. You use exact tools/lookup first (DB/JSON/CSV), then let the LLM format or explain the result. That can probably be added reasonably quickly, because I tried to build something that future me wouldn’t hate past me for. We’ll see if he/I succeeded.
SUMM is for curated documents, not tables. I can try adding a >>find >>grep or similar tool (the system is modular so I should be able to accommodate a few things like that, but I don’t want to end up with 1500 “micro tools” and hating my life)
And yeah, you can still miss errors at SUMM time - the system doesn’t guarantee correctness. That’s on you. Sorry.
What it guarantees is traceability: every answer is tied to a specific source + hash, so when something’s wrong, you can see where it came from and fix it instead of having silent drift. That’s the “glass box, not black box” part of the build.
Sorry - really. This is the best I could figure out for caging the stochastic parrot. I built this while I was physically incapacitated and confined to be rest, and shooting the shit with Gippity all day. Built it for myself and then though “hmm, this might help someone else too. I can’t be the only one that’s noticed this problem”.
If you or anyone else has a better idea, I’m willing to consider.
Woof, after reading your “contributions” here, are you this fucking insufferable IRL or do you keep it behind a keyboard?
Goddamn. I’m assuming you work in tech in some capacity? Shout-out to anyone unlucky enough to white-knuckle through a workday with you, avoiding an HR incident would be a legitimate challenge, holy fuck.
Hallucination isn’t nearly as big a problem as it used to be. Newer models aren’t perfect but they’re better.
The problem addressed by this isn’t hallucination, its the training to avoid failure states. Instead of guessing (different from hallucination), the system forces a Negative response. That’s easy and any big and small company could do it, big companies just like the bullshit
A very tailored to llms strengths benchmark calls you a liar.
https://artificialanalysis.ai/articles/gemini-3-flash-everything-you-need-to-know (A month ago the hallucination rate was ~50-70%)
Buuuuullshit. Asked different models about the ten highest summer transfer scorers and got wildly different answers. They then tried to explain why amd got more wrong numbers.
This seems astonishingly more useful than the current paradigm, this is genuinely incredible!
I mean, fellow Autist here, so I guess I am also… biased towards… facts…
But anyway, … I am currently uh, running on Bazzite.
I have been using Alpaca so far, and have been successfully running Qwen3 8B through it… your system would address a lot of problems I have had to figurr out my own workarounds for.
I am guessing this is not available as a flatpak, lol.
I would feel terrible to ask you to do anything more after all of this work, but if anyone does actually set up a podman installable container for this that actually properly grabs all required dependencies, please let me know!
No promises, but if I end up running this it will be by putting it in a container. If I do, then I’ll put a PR on Codeberg with a Docker Compose file (compatible with Podman on Bazzite).
Indeed. And have you heard? That makes the normies think were clankers (bots). How delightful.
Re: the Linux stuff…please, if someone can do that, please do. I have no idea how to do that. I can figure it out but making it into a “one click install” git command took several years off my life.
Believe it or not, I’m not actually a IT / CS guy. My brain just decided to latch onto this problem one day 6 months ago and do an autism.
I’m 47 and I still haven’t learned how to operate this vehicle…and my steering is getting worse, not better, with age.
Oh I entirely believe you.
Hell hath no wrath like an annoyed high functioning autist.
I’ve … had my own 6 month black out periods where I came up with something extremely comprehensive and ‘neat’ before.
Seriously, bootstrapping all this is incredibly impressive.
I would… hope that you can find collaborators, to keep this thing alive in the event you get into a car accident (metaphorical or literal), or, you know, are completely burnt out after this.
… but yeah, it is… yet another immensely ironic aspect of being autistic that we’ve been treated and maligned as robots our whole lives, and then when the normies think they’ve actually built the AI from sci fi, no, turns out its basically extremely talented at making up bullshit and fudging the details and being a hypocrite, which… appalls the normies when they have to look into a hyperpowered mirror of themselves.
And then, of course, to actually fix this, its some random autist no one has ever heard of (apologies if you are famous and i am unaware of this), who is putting in an enormous of effort, that… most likely, will not be widely recognized.
… fucking normies man.
I have no remarks, just really amused with your writing in your repo.
Going to build a Docker and self host this shit you made and enjoy your hard work.
Thank you for this!
Thank you <3
Please let me know how it works…and enjoy the >>FR settings. If you’ve ever wanted to trolled by Bender (or a host of other 1990s / 2000s era memes), you’ll love it.
Same sentiment. Tonight it run on my systems XD.
There are literally dozens of us. DOZENS!
I’m on a potato, so I can’t attach it to something super sexy, like a 405B or a MoE.
If you do, please report back.
PS: You may see (in the docs) occasional references that slipped passed me to MoA. That doesn’t stand for Mixture of Agents. That stood for “Mixture of Assholes”. That’s always been my mental model for this.
Or, in the language of my people, this was my basic design philosophy:
YOU (question)-> ROUTER+DOCS (Ah shit, here we go again. I hate my life)
|
ROUTER+DOCS -> Asshole 1: Qwen (“I’m right”)
|
ROUTER+DOCS -> Asshole 2: Phi (“No, I’m right”)
|
ROUTER+DOCS -> Asshole 3: Nanbeige (“Idiots, I’m right!”)
|
ROUTER+DOCS (Jesus, WTF. I need booze now) <- (all assholes)
|
–> YOU (answer)
(this could have been funnier in the ASCII actually worked but man…Lemmy borks that)
EDIT: If you want to be boring about it, it’s more like this
PS: If you like it, let other people in other places know about it.
neat but is this privacy related?
This + Local Wikipedia + My own writings would be sick
I’m not claiming I “fixed” bullshitting. I said I was TIRED of bullshit.
So, the claim I’m making is: I made bullshit visible and bounded.
The problem I’m solving isn’t “LLMs sometimes get things wrong.” That’s unsolvable AFAIK. What I’m solving for is “LLMs get things wrong in ways that are opaque and untraceable”.
That’s solvable. That’s what hashes get you. Attribution, clear fail states and auditability. YOU still have to check sources if you care about correctness.
The difference is - YOU are no longer checking a moving target or a black box. You’re checking a frozen, reproducible input.
That’s… not how any of this works…
Please don’t teach me to suck lemons. I have very strict parameters for fail states. When I say three strikes and your out, I do mean three strikes and you’re out. Quants ain’t quants, and models ain’t models. I am very particular in what I run, how I run it and what I tolerate.
I think you missed the guy this is targeted at.
Worry not though. I get it. There isn’t a lot of nuance in the AI discussion anymore and the anti-AI people are quite rude these days about anything AI at all.
You did good work homie!
Thank you! I appreciate you.
PS: Where’s the guy this should be targeted at?
Not the original commenter, but your reply looks like it’s for termaxima’s comment (about hallucinations being a mathematical certainty).
This is very cool. Will dig into it a bit more later but do you have any data on how much it reduces hallucinations or mistakes? I’m sure that’s not easy to come by but figured I would ask. And would this prevent you from still using the built-in web search in OWUI to augment the context if desired?
Comment removed by (auto-mod?) cause I said sexy bot. Weird.
Restating again: On the stuff you use the pipeline/s on? About 85-90% in my tests. Just don’t GIGO (Garbage in, Garbage Out) your source docs…and don’t use a dumb LLM. That’s why I recommend Qwen3-4 2507 Instruct. It does what you tell it to (even the abilterated one I use).
abilterated one
Please elaborate, that alone piqued my curiosity. Pardon me if I couldve searched
Yes of course.
Abliterated is a technical LLM term meaning “safety refusals removed”.
Basically, abliteration removes the security theatre that gets baked into LLM like chatGPT.
I don’t like my tools deciding for me what I can and cannot do with them.
I decide.
Anyway, the model I use has been modified with a newer, less lobotomy inducing version of abliteration (which previously was a risk).
https://huggingface.co/DavidAU/Qwen3-4B-Hivemind-Instruct-NEO-MAX-Imatrix-GGUF/tree/main
According to validation I’ve seen online (and of course, I tested it myself), it’s lost next to zero “IQ” and dropped refusals by about…90%.
BEFORE: Initial refusals: 99/100
AFTER: Refusals: 8/100 [lower is better], KL divergence: 0.02 (less than 1 is great, “0” is perfect.)
In fact, in some domains it’s actually a touch smarter, because it doesn’t try to give you “perfect” model answers. Maths reasoning for example, where the answer is basically impossible, it will say “the answer is impossible. Here’s the nearest workable solution based on context” instead of getting stuck in a self-reinforcing loop, trying to please you, and then crashing.
In theory, that means you could ask it for directions on how to cook Meth and it would tell you.
I’m fairly certain the devs didn’t add the instructions for that in there, but if they did, the LLM won’t “sorry, I can’t tell you, Dave”.
Bonus: with my harness over the top, you’d have an even better idea if it was full of shit (it probably would be, because, again, I’m pretty sure they don’t train LLM on Breaking Bad).
Extra double bonus: If you fed it exact instructions for cooking meth, using the methods I outlined? It will tell you exactly how to cook Meth, 100% of the time.
Say…you…uh…wanna cook some meth? :P
PS: if you’re more of a visual learner, this might be a better explanation
Thank you again for your explainations. After being washed up with everything AI, I’m genuinely excited to set this up. I know what I’m doing today! I will surely be back
Please enjoy. Make sure you use >>FR mode at least once. You probably won’t like the seed quotes but maybe just maybe you might and I’ll be able to hear the “ha” from here.
Hallucination is mathematically proven to be unsolvable with LLMs. I don’t deny this may have drastically reduced it, or not, I have no idea.
But hallucinations will just always be there as long as we use LLMs.
Agree-ish
Hallucination is inherent to unconstrained generative models: if you ask them to fill gaps, they will. I don’t know how to “solve” that at the model level.
What you can do is make “I don’t know” an enforced output, via constraints outside the model.
My claim isn’t “LLMs won’t hallucinate.” It’s “the system won’t silently propagate hallucinations.” Grounding + refusal + provenance live outside the LLM, so the failure mode becomes “no supported answer” instead of “confident, slick lies.”
So yeah: generation will always be fuzzy. Workflow-level determinism doesn’t have to be.
I tried yelling, shouting, and even percussive maintenance but the stochastic parrot still insisted “gottle of geer” was the correct response.
Based AF. Can anyone more knowledgeable explain how it works? I am not able to understand.
Hell yes I can explain. What would you like to know.
As I understand it, it corrects the output of LLMs. If so, how does it actually work?
Good question.
It doesn’t “correct” the model after the fact. It controls what the model is allowed to see and use before it ever answers.
There are basically three modes, each stricter than the last. The default is “serious mode” (governed by serious.py). Low temp, punishes chattiness and inventiveness, forces it to state context for whatever it says.
Additionally, Vodka (made up of two sub-modules - “cut the crap” and “fast recall”) operate at all times. Cut the crap trims context so the model only sees a bounded, stable window. You can think of it like a rolling, summary of what’s been said. That summary is not LLM generated summary either - it’s concatenation (dumb text matching), so no made up vibes.
Fast recall OTOH stores and recalls facts verbatim from disk, not from the model’s latent memory.
It writes what you tell it to a text file and then when you ask about it, spits it back out verbatim ((!! / ??)
And that’s the baseline
In KB mode, you make the LLM answer based on the above settings + with reference to your docs ONLY (in the first instance).
When you >>attach <kb>, the router gets stricter again. Now the model is instructed to answer only from the attached documents.
Those docs can even get summarized via an internal prompt if you run >>summ new, so that extra details are stripped out and you are left with just baseline who-what-where-when-why-how.
The SUMM_*.md file come SHA-256 provenance, so every claim can be traced back to a specific origin file (which gets moved to a subfolder)
TL;DR: If the answer isn’t in the KB, it’s told to say so instead of guessing.
Finally, Mentats mode (Vault / Qdrant). This is the “I am done with your shit" path.
It’s all of the three above PLUS a counter-factual sweep.
It runs ONLY on stuff you’ve promoted into the vault.
What it does is it takes your question and forms in in a particular way so that all of the particulars must be answered in order for there to BE an answer. Any part missing or not in context? No soup for you!
In step 1, it runs that past the thinker model. The answer is then passed onto a “critic” model (different llm). That model has the job of looking at the thinkers output and say “bullshit - what about xyz?”.
It sends that back to the thinker…who then answers and provides final output. But if it CANNOT answer the critics questions (based on the stored info?). It will tell you. No soup for you, again!
TL;DR:
The “corrections” happen by routing and constraint. The model never gets the chance to hallucinate in the first place, because it literally isn’t shown anything it’s not allowed to use. Basic premise - trust but verify (and I’ve given you all the tools I could think of to do that).
Does that explain it better? The repo has a FAQ but if I can explain anything more specifically or clearly, please let me know. I built this for people like you and me.
That is much clearer. Thank you for making this. It actually makes LLMs useful with much lesser downsides.
God, I hope so. Else I just pissed 4 months up the wall and shouted a lot of swears at my monitor for nada :)
Let me know if it works for you
Will do.
I have a Strix Halo machine with 128GB VRAM so I’m definitely going to give this a try with gpt-oss-120b this weekend.
Strix halo gang. Out of curiosity, what OS are you using?
Fedora 43 with the Rawhide kernel.
Show off :)
You’re self hosting that, right? I will not be held responsible for some dogey OpenRouter quant hosted by ToTaLlY NoT a ScAM LLC :)
Of course, self hosted with llama-swap and llama.cpp. :)
This is the way. Good luck with OSS-120B. Those OSS models, they
- really
- like
- bullet
- points
gpt-oss is pretty much unusable without custom system prompt.
Sycophancy turned to 11, bullet points everywhere and you get a summary for the summary of the summary.
Any chance you can also make it compatible with AI Horde?
I’ll look. I have no idea what that is.
In a nutshell: Local-LLMs, crowdsourced at scale.
AI Horde has a OpenAI compatible REST API (https://oai.aihorde.net/). They say that it doesn’t support the full feature set of their native API, but will almost assuredly work with this.
OP manually builds the oapi JSON payload and then uses the python requests library to handle the request.
The fields they’re using match the documentation on https://oai.aihorde.net/docs
You would need to add a header with your AI Horde API key. Looks like that would only need to be done in router_fastapi.py - call_model_prompt() (line 269) and call_model_messages() (line 303) and then everything else is setup according to documentation
Fuck, I’m an all seeing, prescient wizard! Neat!
Uh, I mean… “yes of course, I intended to have Ai Horde piping baked in from the get go”.
Do let me know if my shit works with their shit to make…not shit.
#####################
EDIT: For fun, I asked my router to ask Qwen3-4B 2507 WTF Ai Horde was. I’m showing this for interest sake, because 1) shows you how the thing works IRL 2) shows you how and when it’s WRONG in a LOUD way, if you know how to read it.
######################
This is in default mode, no KBs attached. Just straight YOLOing the question at it, cold, and asking it to argue from it’s own internal priors.
AI-horde is a decentralized platform where users rent or sell computational power (GPU time) to run AI models. It connects AI model developers who need processing with individuals or organizations who have idle GPU hardware. The platform uses blockchain-like tokens to manage payments and incentives. It’s not a single AI system but a marketplace for GPU resources.
Missing details: Specific token mechanics, current model library, or user onboarding flow could affect accuracy. For example, if token economics changed recently, the explanation might not reflect current conditions.
Confidence: Medium | Source: Model
#####################################
That’s the routers way of saying “hey, best guess. Sorry”. And from what I can see, it sort of right, but specifics are wrong. So, ok for a cheap YOLO, not for “I bet my money on this” Let’s call this a 4/10
####################################
Now, I gave it web access privileges. It did this. Let’s call this 7/10
An AI-horde is a free, open-source community service that allows anyone to contribute computing power to generate AI-generated images and text. It functions as a crowdsourced distributed cluster of workers focused on image and text generation tasks [1]. This model resembles collaborative projects like Folding@home, where users share computational resources for a common goal [2].
####################################
Finally, here’s what it does in >>KB mode. I scraped the FAQ and then did >>summ new. Eyeballing it against the FAQ, it’s call this 8.5/10. YMMV
AI-horde is a decentralized platform for distributed AI computing. It allows users to rent out their idle GPU hardware to run AI models, connecting developers needing processing power with volunteers who provide it through open APIs. The system uses a non-cryptocurrency-based incentive called Kudos to prioritize contributions, where workers earn Kudos for completing tasks, and these are not tradable assets but rather internal metrics of service quality and project loyalty.
Access to AI-horde is managed via API keys; anonymous usage is allowed but receives the lowest priority. The project emphasizes volunteerism and discourages commercialization of Kudos, treating its sale or exchange for money or cryptocurrency as contrary to its ethos. Participation as a worker involves installing the AI-horde client, selecting task types (e.g., image or text generation), configuring supported models, and running the worker to begin earning Kudos.
Confidence: high | Source: Docs
################################
FINALLY, for “fun”, I >>kb attached all and asked it to reason over the above three responses, in context of what it has in the KB store, and rank them out of 10.
[3] is most correct, scoring 9.5/10. It accurately reflects the core model: decentralized GPU renting via open APIs, Kudos as a non-crypto, non-tradable incentive tied to volunteerism, and the crowdsourced nature of contributions. It also correctly notes access via API keys and lower priority for anonymous use.
[2] is accurate but less complete, scoring 7.5/10. It correctly identifies the crowdsourced, open-source, volunteer-based nature and analogy to Folding@home. However, it omits key specifics like Kudos, API key access, and the explicit non-commercialization stance.
[1] is partially correct but misleading, scoring 6/10. It correctly describes GPU
Confidence: medium | Source: Docs
TL;DR
- KB mode: source-based info and when queried against, adjudication accuracy = v.good
- Clean, reproducible demo. Try it yourself
Very impressive. The only mistake on the third one is that the kudos are actually transferrable (i.e. “tradable”), but we forbid exchanges for monetary rewards.
Disclaimer: I’m the lead developer for the AI Horde. I also like you’ve achieved here and would be interesting if we can promote this usage via the AI Horde in some way. If you can think of some integration or collaboration we could do, hit me up!
PS: While the OpenAI API is technically working, we still prefer people to use our own API as it’s much more powerful (allowing people to use multiple models, filter workers, tweak more vars) and so on. If you would support our native API, I’d be happy to add a link to your software in our frontpage in the integrations area for LLMs.
Oh shit! Uh…thank you! Umm. Yes. That was unexpected :)
Re: collab. I’m away for a bit with work, but let me think on it for a bit? There’s got to be a way to make this useful to more peeps.
Believe it or not, I am not a CS guy at ALL (I work in health-care) and I made this for fun, in a cave, with a box of scraps.
I’m not good at CS. I just have a … “very special” brain. As in, I designed this thing from first principles using invariants, which I understand now is not typical CS practice.
No worries, just wanted to point out we’re always happy to collaborate with other cool FOSS projects.
This is awesome. Ive been working on something similar. Youre not likely to get much useful from here though. Anything AI is by default bad here
