Show HN: I'm tired of my LLM bullshitting. So I fixed it
Posted by BobbyLLM 2 days ago
As a handsome local AI enjoyer™ you’ve probably noticed one of the big flaws with LLMs:
It lies. Confidently. ALL THE TIME.
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 (eg: OWUI) & backend (llama.cpp + llama-swap). Local-first but it should talk to anything OpenAI-compatible if you point it there (note: experimental so YMMV).
LC is 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 (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 - generates SUMM_.md files 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."
[Confidence: medium | Source: Mixed]
No vibes. Just: here’s what’s in your docs, here’s what’s missing, don't GIGO yourself into stupid.
Then, if you're happy with the summary, you can:
>>move to vault
2. Mentats: proof-or-refusal mode (Vault-only)
Mentats is the “deep think” pipeline against your curated sources.
* no chat history
* no filesystem KBs
* no Vodka
* Vault-only grounding (Qdrant)
It runs a 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. 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
Potato PCs have two classic problems: goldfish memory + context bloat that murders your VRAM.
Vodka fixes both without extra model compute.
* !! 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) and creates a concatenated summary (not LLM) 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 or language are weird, now you know why.
Comments
Comment by dankwizard 1 day ago
Comment by BobbyLLM 1 day ago
Prompts shape style, not epistemics.
What this does is moves the problem out of “please behave” and into hard constraints.
Nice drive by tho.
PS: You understand this is enforced outside the model, right? Or are you here just to try and dunk on someone?
Comment by gogopromptless 1 day ago
Comment by BobbyLLM 1 day ago
Do you mean generally though - how many facts does it extract from a typical document of X length? Or do you mean what my own personal corpus currently is?
Comment by gogopromptless 1 day ago
My crappy personal system is a telegram channel "Today I Learned" that grows at roughly 5 messages per day. The search is obviously much worse than your tool.
The design of no rewrites for facts storage was an approach I have not seen from any of my friends' agent memory setups, so that difference struck my curiosity for how well its working in practice.
Comment by BobbyLLM 1 day ago
If you're talking about the "Fastrecall" part of it, it's clamped to around 3000 entries. Beyond that...I mean...it's just a JSON file. How much do you hate yourself? lol :)
I use mine as intended (well, by me) as a short term memory store with a TTL of 14 days. Right now it has...350ish stored facts with different expiration dates. You can play around with the settings and max resets (touches), but if ctx > max_items, janitor rolls in and the oldest get pruned anyway.
If you're talking about the chat bloat control part (cut the crap), I set mine to keep the last 2 user/assistant pairs and a soft cap of 2000 characters. That middle part obviously decays during the chat. I like to keep my chatty fast.
If you're asking about the KB (attach) system: generally speaking, the SUMM function seems to generate a summary of upto about 1500 characters. That's what...250 words? Short and sweet...because again, potato pc + smart grep = don't be stupid.
In theory, the KB thing is limited by you SSD and pain tolerance. I have about 1500 items I query against and it's still in the sub-secondish range.
Mentats / vault? Scales via Qdrant.
Does that answer it? None of this shit is enterprise grade; it's purely for personal use.
Comment by reify 1 day ago
"bro I spent all weekend in claud code. Its incredible"
"Oh nice, what did you build?"
"Dude my set up is crazy. I've got all the vercel skills. plus custom hooks for every project"
"Sick, what are you building?"
"My set up is so optimized. I'm using, like 5 instances at once"
"Hello bro, stop playing with your dick, answer my questions bro!"
Comment by BobbyLLM 1 day ago
Also, Claude said he likes me better than you and he'd know because he's French and old soul and shit.
PS: The concrete thing is a local control layer that makes my LLM shut up unless it can cite my docs. Plus I can run in on my p.o.s P1000 and not line NVIDIA shareholders pockets by buying more VRAM.
Repo’s over there if you want to kick the tires.
Comment by BobbyLLM 1 day ago
I asked my router to ask: "WTF is AI Horde?" and I’m posting the full outputs because this shows (a) how it behaves in different modes, and (b) how and when it’s wrong in a loud, readable way.
1) Default / YOLO mode (no KB attached). Cold question. Model argues from its 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
My take: Meh. Sort of. Wrong specifics. Fine for a cheap guess, not fine for “I’m betting money on this.” Note: it clearly identifies "meh, dunno" parts. Call it a 4/10 for factual correctness, with a clear shoulder shrug.
2) Web-enabled mode
Same question, but allowed to pull from the web. Now it sez -
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].
My take: Better. Still missing details, but broadly right. Let's call it ~7/10.
3) KB mode (FAQ scraped - >>summ new - query)
Here the model is constrained to just what’s in the attached KB summaries.
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
My take: That seems pretty solid well source-grounded. Call it 8/10 based on docs.
Finally -
4) For fun, I then asked it to rank the three prior responses against the KB facts. Here's what it said -
[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
Point being: ungrounded self-eval is vibes. Grounded evaluation becomes boring and useful. Also, Qwen is an easy marker.
TL;DR
YOLO mode: plausible guesses, sometimes confidently wrong. Same old, same old.
Web mode: better, still imperfect.
KB mode: source-based answers with explicit limits/refusals
KB-grounded ranking: reproducible, source-anchored adjudication
Repeatable demo. Attach good model and try it yourself.