llama-conductor
codeberg.org
Route workflows, not models. Glass-box, not black-box. Squash LLM nonsense.

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).

Not a model, not a UI, not magic voodoo.

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 — 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. 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:

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.

      • SuspciousCarrot78@lemmy.worldOP
        18·
        10 hours ago

        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.