I’ll cop to that. At a high level it is “tool calling + RAG + guardrails”.

Ok.

But that’s sort of the point: boring plumbing that turns LLMs from improv actors into constrained components.

Addressing your points directly as I understand them -

1) Doesn’t prevent lying

If you mean “LLMs can still hallucinate in general”, yes. No argument. I curtailed them as much as I could with what I could.

But llama-conductor isn’t trying to solve “AI truth” as a metaphysical problem. It’s trying to solve a practical one:

In Mentats mode, the model is not allowed to answer from its own priors or chat history. It only gets a facts block from the Vault. No facts → refusal (not “best effort guess").

That doesn’t make the LLM truthful. It makes it incapable of inventing unseen facts in that mode unless it violates constraints - and then you can audit it because you can see exactly what it was fed and what it output.

So it’s not “solving lying,” it’s reducing the surface area where lying can happen. And making violations obvious.

2) Wouldn’t a normal search algorithm be better?

I don’t know. Would it? Maybe. If all you want is “search my docs,” then yes: use ripgrep + a UI. That’s lighter and more portable.

The niche here is when you want search + synthesis + policy:

• bounded context (so the system doesn’t slow down / OOM after long chats)
• deterministic short-term memory (JSON on disk, not “model remembers")
• staged KB pipeline (raw docs → summaries with provenance → promote to Vault)
• refusal-capable “deep think" mode for high-stakes questions

I think an algo or plain search engine can do wonders.

It doesn’t give you a consistent behavioral contract across chat, memory, and retrieval.

3) “Everything looks like a nail”

Maybe. But the nail I’m hitting is: “I want local LLMs to shut up when they don’t know, and show receipts when they do.”

That’s a perfectly cromulent nail to hit.

If you don’t want an LLM in the loop at all, you’re right - don’t use this.

If you do want one, this is me trying to make it behave like infrastructure instead of “vibes”.

Now let’s see Paul Allen’s code :P

if you self host, you belong. That’s non-negotiable

Ah. So -

First prize: picture of you

Second prize: two pictures

?

:P

It’s a very nice concept to be able to dump a shitload of knowledge into a folder, look at “processing…” for a few seconds, then ask questions to get exactly what you’re looking for out of it rather than having to go digging through the mound of information and without having to worry that the computer just threw in a few made up facts for giggles. The idea is that the dumping happens over time mostly, allowing you to quickly find buried information from years ago with a few relevant queries.

One thing I’d do with this is dump all my emails into it, from across all my accounts. That might save me having to search keywords in 8+ accounts over 4-5 different platforms every so often…

It also might have been useful in a lawsuit I prosecuted a few years ago. Instead of going through two years of encrypted messages by hand to pull out relevant excerpts with context, I could have exported the lot and just asked for the information. If it worked it could have saved me months (I spent a few hours after work every night screenshotting, dating the screenshots in chronological order, then I’ve that was done I kept a spreadsheet that I filled with relevant excerpts and links to the screenshots, by reading every single screenshot… it was a lot).

Correct. Curate your sources :)

I can’t LoRa stupid out of a model…but I can do this. If your model is at all obedient and non-stupid, and reasons from good sources, it will do well with the harness.

Would you like to see the benchmarks for the models I recommend in the “minimum reccs” section? They are very strong…and not chosen at random.

Like the router, I bring receipts :)

Oh it can try…but you can see its brain. That’s the glass box part of this. You can LITERALLY see why it says what it says, when it says it. And, because it provides references, you can go and check them manually if you wish.

Additionally (and this is the neat part): the router actually operates outside of the jurisdiction of your LLM. Like, the LLM can only ask it questions. It can’t affect the routers (deterministic) operation. The router gives no shits about your LLM.

Sometimes, the LLM might like to give you some vibes about things. Eg: IF YOU SHOUT AT IT LIKE THIS, the memory module of the router activates and stores that as a memory (because I figured, if you’re shouting at the llm, it’s probably important enough in the short term. That or your super pissed).

The llm may “vibe” a bit (depending on the temp, seed, top_k etc), but 100/100, ALL CAPS >8 WORDS = store that shit into facts.json

Example:

User: MY DENTIST APPOINTMENT IS 2:30PM ON SATURDAY THE 18TH.

LLM: Gosh, I love dentists! They soooo dreamy! <----PS: there’s no fucking way your LLM is saying this, ever, especially with the settings I cooked into the router. But anywayz

[later]

USER: ?? When is my dentist appointment again

LLM: The user’s dentist appointment is at 2:30 PM on Saturday, the 18th. The stored notes confirm this time and date, with TTL 4 and one touch count. No additional details (e.g., clinic, procedure) are provided in the notes.

Confidence: high | Source: Stored notes

Yes, I made your LLM autistic. You’re welcome

Built by an autist to give your LLM autism. No Tylenol needed.

On the AI slop image. I sez to shitGPT, I sez “Yo, make me ZARDOZ but you know, cute and chibi like”

Enjoy the nightmare fuel

[ZARDOZ HATH SPOKEN]

No. Notes apps store your text. This is a control panel for how a LLM reasons over them.

“Extra steps” lmao

Not intentionally :)

Yes! And, for fun, I forced Qwen to answer your question, after querying its own fucking brain (M-E-T-A).

The below is AI generated, using the router, with >>Fun mode (not >>FR…that’s my special gremlin mode). >>f suspends some of the terseness. Let me know if you want to see the non FR version

Yep — with your current code, here’s the literal answer to “what can >>attach ingest, can it handle the whole Python docs, and how does retrieval work?” (based on the files you uploaded).

What >>attach actually does (and does not do)

>>attach <kb> doesn’t ingest anything. It just registers a KB name into state.attached_kbs so later turns know which KB folders are “active.”

The router’s own docstring is blunt about the intended contract:

• “KBs are filesystem folders containing SUMM_*.md files.”
• “/serious answers use filesystem KB retrieval ONLY from attached KBs.”

So the “ingest” step is really:

1. drop raw docs in the KB folder
2. run >>summ new to generate SUMM_*.md
3. then attach the KB (or keep it attached) and query it

What kinds of files can be summarized (and therefore used via attach)?

In the router’s filesystem SUMM pipeline, _SUPPORTED_RAW_EXTS = {".md", ".txt", ".pdf", ".html", ".htm"}

• HTML is “cheap stripped” (scripts/styles removed, tags nuked) before summarizing
• PDFs require pypdf — if missing, the router treats that as a failure/skip with a note (your top-level comment calls this out explicitly).
• There’s also an explicit guard to truncate huge inputs before sending to the model (default summ.max_input_chars = 120_000).

When a SUMM is created, it writes a provenance header including source_rel_path and source_sha256, then moves the original into /original/.

So: you do not need “minimalistic plain-text statements.” You can feed it normal docs (md/txt/html/pdf) and it will produce SUMMs that become queryable.

“If I dropped the entire Python docs in there…”

Yes, it will produce something usable, because Python docs are mostly HTML and you explicitly support .html/.htm with stripping.

But there are two practical gotchas in your implementation:

1. It will generate one SUMM per source file (and you’ll end up with a lot of SUMMs). summ_new_in_kb() walks the KB tree, skips /original/, and summarizes every supported raw doc that doesn’t already have a corresponding SUMM_*.md.
2. The SUMM prompt structure matters. Your shipped SUMM.md template is oriented around “overview / key ideas / steps / tradeoffs / pitfalls” rather than “API reference / signatures / parameters.” So it’ll work better for conceptual docs than for “tell me the exact signature of pathlib.Path.glob”.

If you want Python-docs-as-KB to be strong, the move is: tweak the SUMM prompt to extract API sections + signatures + examples, not just “Key Ideas.”

How retrieval works in “attached KB” mode

This is the most important bit: filesystem KB retrieval is not vector embeddings. It’s a lightweight lexical scorer over SUMM markdown blocks.

Concretely:

• It searches only SUMM_*.md under attached KB roots, explicitly excluding /original/.
• It splits each SUMM into blocks (paragraphs / headings), caps blocks around ~1200 chars, then scores blocks by token overlap with the query.
• Scoring has a few boosts: headings matching query tokens, and a small bonus for code fences.
• Then it returns the top hits (defaults like top_k=8, max_blocks_per_file=3, max_chars=2400).

So attached-KB mode is basically: pre-summarize once → then do fast “smart grep” over summaries.

That’s why it’s potato-friendly: you’re not embedding every doc on every query; SUMMs are preprocessed once.

And “Vault / Mentats” retrieval is different

When you >>move to vault, you’re no longer doing filesystem lexical search. You promote SUMMs into Qdrant and retrieve top-k chunks with embeddings + rerank (as configured in README / config).

(Also worth noting: per your ingest module, the intended architecture is explicitly “filesystem SUMMs for day-to-day; Vault/Qdrant only for promoted SUMMs.” )

One-liner you can paste as a Lemmy answer

>>attach doesn’t ingest raw docs — it just activates a KB folder. The KB mode queries SUMM_*.md files (generated via >>summ new) using a lightweight lexical block scorer (token overlap + heading boosts), not embeddings. You can drop the whole Python HTML docs in a KB and it’ll work, but you’ll want a SUMM prompt tuned for API/reference detail or the summaries will smooth out the stuff you actually care about. If you promote SUMMs to the Vault (>>move to vault), then retrieval switches to Qdrant vector search + reranking.

###########################

Is gut, ja?

PS: You can tweak the SUMM.md yourself because it’s not baked into the code or nuthin. Hack away :)

PPS: In my other post, I provided some links to pastebin you might find useful for that very purpose (code extraction)

https://lemmy.world/post/41992574/21702596

(see the ##My RAG post)

This is a quote from Deming, one of the fathers of modern data analysis. It basically means “I don’t trust you. You’re not god. Provide citations or retract your statement”

Exactly what it says, what’s confusing about it? The problem with LLMs is that they bullshit; the problem with LLMs bullshitting is that you have to check everything they say, they are not trustworthy - severely limiting the utility. So, don’t trust LLMs. God meanwhile is assumed to be the paragon of honesty, and thus completely trustworthy, so would be the only entity to be trusted implicitly.

This works even if you don’t believe in god, like me: trust nobody, everyone must bring data if they want me to trust what they say.

https://en.wikipedia.org/wiki/In_God_We_Trust%3A_All_Others_Pay_Cash

“Trust” should only be for the supernatural, the axiom of reality, and your interpersonal relationships.

Everything else can be proven.