CAEL is an assistant I build that uses no neural network when it runs. It is a set of small engines: counted language models, cellular automata, symbolic solvers for logic and arithmetic, a bounded Python interpreter and local Wikipedia retrieval. Version 0.3.2 adds one more: when CAEL doesn't know an answer, it searches the web.
How the web search works
- It searches Brave Search (free monthly tier), then a SearxNG instance, DuckDuckGo Instant Answers and Wikipedia. When one fails or runs out of free searches, the next one is used.
- It reads the top pages as clean text through Jina Reader.
- It scores each sentence by the question's key words and the kind of answer asked for (a year, a name, a number), and rejects sentences that only name the subject. "Scotland is a country..." does not answer "What is the national animal of Scotland?".
- It says "confirmed" only when two or more different websites give the same answer. Otherwise it says it is not sure and links the one source so you can check it.
What it changes
I wrote new questions for the test, so none come from a benchmark:
| No internet | Wikipedia only | With web search | |
|---|---|---|---|
| 44 questions (used to find bugs) | 23% | 59% | 86% |
| 40 questions written after the fixes | 10% | 35% | 73% |
Web search does nothing for reasoning. On theory-of-mind stories ("where will Anna look for the ball?") CAEL scores about 24%, with or without the internet. A web answer also takes 5-25 seconds, against a fraction of a second for what CAEL already knows.
Try it
git clone https://github.com/Vovala14/CAEL.git
cd CAEL
python -m pip install -r requirements-unified.txt
python -m cael_unified --download
python -m cael_unified --ui
Code: https://github.com/Vovala14/CAEL
Model files: https://huggingface.co/vynly/CAEL-0.3.2
Benchmark: https://huggingface.co/spaces/vynly/non-neural-ai-leaderboard
Free for personal and research use under the CAEL Non-Commercial License.
This post was written with help from an AI assistant.
Top comments (0)