You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Commit 231c688
Browse filesBrowse the repository at this point in the historyBrowse files
docs: appealing README with badges and AI-papers demo familiar to peer users
- hero, badges, workflow mapping from LLM Wiki/Reor/Khoj, providers, comparison
- demo.svg now shows reading AI papers (Transformer, BERT, RAG) from the real core
- data/demo/papers sample set; make start opens it by default
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
**A native GNOME knowledge workstation that turns raw material into an evolving knowledge base.**
5
+
<h1align="center">gknowledge</h1>
4
6
5
-
`gknowledge` is a desktop application for GNOME that helps users collect documents, notes, links, and research material, then incrementally transform that corpus into a structured knowledge space with the support of LLMs.
7
+
<palign="center">
8
+
<b>Drop in your documents. Get a linked wiki that an LLM keeps up to date.</b><br>
9
+
A native GNOME app. Offline by default, plain Markdown on disk, Obsidian-compatible.<br>
10
+
Bring your own brain: a local model, any OpenAI-compatible endpoint, or your own ChatGPT plan.
11
+
</p>
6
12
7
-
The goal is not to build yet another chat wrapper.
8
-
The goal is to build a serious native app for Linux desktops that makes knowledge curation, synthesis, and retrieval feel like a first-class operating system capability.
<sub>The animation is generated from the real core on the chess sample documents in `data/demo/chess` (offline engine, no API key). Regenerate it with `make demo-svg`.</sub>
36
+
<palign="center">
37
+
<imgsrc="docs/demo.svg"alt="gknowledge in action: drop in AI research papers, get a linked wiki, ask a question and get an answer with citations"width="880">
38
+
</p>
39
+
40
+
<palign="center"><sub>
41
+
Real output of the offline engine on six AI papers (<code>data/demo/papers</code>), no API key.
42
+
Regenerate with <code>make demo-svg</code>.
43
+
</sub></p>
44
+
45
+
## Why gknowledge
46
+
47
+
If you have used [LLM Wiki](https://github.com/nashsu/llm_wiki), [Tolaria](https://github.com/refactoringhq/tolaria), [OpenKnowledge](https://github.com/inkeep/open-knowledge), [Reor](https://github.com/reorproject/reor) or [Khoj](https://github.com/khoj-ai/khoj), you already know the loop: **ingest sources, let the model build and maintain a wiki, then ask questions against it.** gknowledge does the same thing as a real GNOME application.
48
+
49
+
-**Native, not a web view.** GTK 4 and Libadwaita, built to follow the GNOME HIG. No Electron, no Tauri, no browser engine.
50
+
-**Works with no account.** A built-in heuristic engine ingests and answers offline. Add a model when you want better synthesis.
51
+
-**Your data stays yours.** Markdown with YAML frontmatter plus a readable SQLite index. Open the folder in Obsidian or any editor.
52
+
-**Cited, reviewable, incremental.** Every generated page records its sources, answers cite pages, and a review queue keeps you in control.
53
+
54
+
## The workflow you already know
55
+
56
+
| You know this from… | What it is | In gknowledge |
57
+
|---|---|---|
58
+
| Karpathy's LLM Wiki, LLM Wiki |**Ingest**: read a new source, update the pages it touches | Done: two-step extraction with provenance |
59
+
| Reor, Khoj |**Ask**: questions answered from your notes, with citations | Done: lexical search plus graph expansion, citations, saved answers |
make start # opens a demo project with six AI papers already ingested
75
+
```
76
+
77
+
`make start` also restarts the app whenever you edit the source, which makes it the fastest way to hack on gknowledge. Install it for daily use with `make install`.
Attention Is All You Need is a 2017 paper by Ashish Vaswani and seven co-authors at Google Brain and Google Research. It introduced the Transformer, a neural network architecture based entirely on self-attention that dispenses with recurrence and convolutions.
4
+
5
+
The Transformer uses multi-head attention, so the model can look at different positions of a sequence in parallel, and adds positional encodings to keep word order. Because it avoids sequential computation, it trains much faster than recurrent networks.
6
+
7
+
The model reached 28.4 BLEU on the WMT 2014 English-to-German translation task, a new state of the art, after training for 3.5 days on eight P100 GPUs. The paper was presented at NeurIPS 2017 and became the foundation of BERT, GPT-3 and almost every modern large language model.
BERT, short for Bidirectional Encoder Representations from Transformers, was introduced in 2018 by Jacob Devlin, Ming-Wei Chang, Kenton Lee and Kristina Toutanova at Google AI Language. The paper was published at NAACL 2019.
4
+
5
+
BERT pre-trains a Transformer encoder on unlabelled text with two objectives: a masked language model, which hides some words and asks the model to predict them, and next sentence prediction. Because it reads context from both directions, it builds richer representations than left-to-right models.
6
+
7
+
BERT-base has 110 million parameters and BERT-large has 340 million. After fine-tuning it set a new state of the art on eleven natural language processing tasks, including the GLUE benchmark and the SQuAD question answering dataset.
Chain-of-Thought Prompting Elicits Reasoning in Large Language Models is a 2022 paper by Jason Wei and colleagues at Google Brain, presented at NeurIPS 2022.
4
+
5
+
Instead of showing only question and answer pairs, the prompt includes worked examples with the intermediate reasoning steps written out. Large models then produce their own step-by-step reasoning before giving the final answer. This sharply improves arithmetic, commonsense and symbolic reasoning.
6
+
7
+
With eight chain-of-thought examples, PaLM 540B reached state-of-the-art accuracy on GSM8K, a benchmark of grade-school math word problems, surpassing fine-tuned GPT-3 with a verifier. The ability emerges only at large scale, roughly around one hundred billion parameters, and does not help small models.
Language Models are Few-Shot Learners is the 2020 paper by Tom Brown and colleagues at OpenAI that introduced GPT-3. The model has 175 billion parameters, more than ten times larger than any earlier non-sparse language model.
4
+
5
+
The key finding is in-context learning. Given only a few examples written in the prompt, GPT-3 performs translation, question answering and arithmetic without any gradient update or fine-tuning. Performance improves smoothly with model size, and few-shot results approached those of fine-tuned systems on many benchmarks.
6
+
7
+
GPT-3 was trained on roughly 300 billion tokens of filtered Common Crawl, books and Wikipedia. The paper also discussed risks such as bias and misuse for generating misleading text.
0 commit comments