TooCas

TooCas / SMELT

Public

SMELT: Schema-Aware Markdown Compilation for Efficient Local Token Inference — 95% token reduction on query-conditioned retrieval

19
0
100% credibility
Found Apr 03, 2026 at 19 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Python
AI Summary

SMELT is a tool that transforms static markdown notes used in AI agent workspaces into efficient, query-specific formats to minimize redundant processing during conversations.

How It Works

1
🔍 Discover SMELT

You hear about a helpful tool that makes your AI helper read your notes much faster by only focusing on what's needed for each question.

2
💻 Get it ready

You download the tool to your computer and set it up in a few moments so it's good to go.

3
📁 Pick your notes

You select the important note files from your AI setup that get shared over and over.

4
⚡ Make smart packs

You create compact versions of your notes that automatically pull out just the right info for any question.

5
❓ Ask about something

You type a simple question like 'Who is Kane?' and the tool finds exactly what's relevant.

6
📤 See the savings

You get back a short, perfect snippet of info instead of your whole big file, making everything quicker.

🎉 Speedy AI magic

Your AI chats now zoom along faster, use less power, and feel smarter without missing a beat.

Sign up to see the full architecture

5 more

Sign Up Free

Star Growth

See how this repo grew from 19 to 19 stars Sign Up Free
Repurpose This Repo

Repurpose is a Pro feature

Generate ready-to-use prompts for X threads, LinkedIn posts, blog posts, YouTube scripts, and more -- with full repo context baked in.

Unlock Repurpose
AI-Generated Review

What is SMELT?

SMELT is a Python tool for schema-aware markdown compilation that slashes token usage in local LLM inference by up to 95% through query-conditioned retrieval. It targets agent frameworks dumping static workspace files—like user profiles or memory docs—into every model call, compiling them into a dense, relevant subset per query. Run it via CLI: compile your markdown files once, then query for efficient retrieval without reprocessing everything.

Why is it gaining traction?

Unlike naive JSON dumps or heading strippers that barely dent tokens, SMELT's layered approach—lossless archival, semantic flattening, and smart retrieval—delivers real savings, like 94.7% on targeted questions, while preserving context. Developers notice faster startup times (6% TTFT drop) and lower latency on local setups, with benchmarks using actual tokenizers like Qwen's. It's a practical fix for smelting bloated contexts down to essentials, much like a Minecraft smelter refining ore.

Who should use this?

Local inference runners on Apple Silicon or similar hardware building agentic apps with frameworks like OpenClaw. Indie researchers optimizing multi-file workspaces for production sessions. Python devs handling markdown-heavy RAG pipelines who hate paying for repeated static tokens.

Verdict

Promising for token-hungry local agents, with solid benchmarks and a DOI-backed paper, but at 19 stars and 1.0% credibility, it's early-stage—test thoroughly before production. Pair it with your tokenizer for real wins.

(178 words)

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.