CAEL 0.3.2: internet search when it does not know
CAEL is a hybrid non-neural assistant. It combines counted language models, cellular automata, symbolic reasoning, exact arithmetic, a bounded Python interpreter, local retrieval and hand-written knowledge. It uses no neural network, transformer or external language model when it runs. This is a custom Python runtime, not a Transformers/AutoModel checkpoint.
Download and run
- Get the source: GitHub, Vovala14/CAEL, or CAEL-0.3.2.zip from this repository.
- Install and fetch the checksum-verified 553 MB text-only bundle:
python -m pip install -r requirements-unified.txt
python -m cael_unified --download
python -m cael_unified "Who wrote the opera The Magic Flute?" --mode chat
python -m cael_unified --ui # http://127.0.0.1:8765
On Windows, SETUP.bat and then RUN.bat do the same. For an offline install, put
cael-0.3.2.cael in the source folder's
models/ directory and run python -m cael_unified --download.
What's new in 0.3.2
- Web search fallback. When a factual question has no local answer, CAEL searches the web, reads the pages
and quotes the sentence that answers it.
- Search order: Brave Search (set
BRAVE_API_KEY; free monthly tier), a SearxNG instance (CAEL_SEARXNG_URL), DuckDuckGo Instant Answers, then Wikipedia. When one fails or its free tier runs out, the next one is used. - Pages are read as clean text through Jina Reader, or by a direct fetch that only reaches public addresses.
- Every web answer links its source. "N different websites agree" appears only when two or more websites give the same answer; otherwise CAEL says it is not sure and asks you to check the source.
- A sentence that only names the subject ("Scotland is a country ...") is not taken as the answer.
- "Write a Python script that searches the web for ..." gives the same search as a runnable script.
- Off with
CAEL_WEB_SEARCH=0. With it on, your question is sent to the search services and Jina Reader.
- Search order: Brave Search (set
- More Wikidata facts: currency, boiling and melting points, height, length, opening and release dates, founding dates and creators.
- Multiple choice: a question with A–D options always goes to the multiple-choice engines.
- Year answers from Wikipedia must match every word of the event.
Results (local, maintainer-run)
New questions written for this test (no benchmark items):
| No internet | Wikipedia/Wikidata only | With web search | |
|---|---|---|---|
| Set 1 (44 questions) | 10 (23%) | 26 (59%) | 38 (86%) |
| Set 2, written after the fixes (40) | 4 (10%) | 14 (35%) | 29 (73%) |
How much to trust these numbers:
- Set 1 was used to find the problems fixed in this release, so treat 86% as a development score.
- Set 2 was written after the fixes and is the fairer estimate.
- Answers were graded by keyword; two set-1 answers counted correct were only partial.
- A web answer takes about 5–25 seconds and depends on what the search services return that day.
Web search does not help with theory-of-mind stories (Tiny Theory of Mind: about 24%). On NNAI-Bench, a maintainer run of 0.3.1 scored 68.11%, marked test-adapted on the leaderboard.
Licence and data
- CAEL Non-Commercial License (LICENSE): free for personal and research use. Commercial use needs a separate written licence from the copyright holder (Vovala14).
- CAEL only uses third-party material that allows commercial use. Components derived from CC BY-SA sources (Wikipedia, Stack Exchange, Dolly, Wikibooks, Wikivoyage) remain under CC BY-SA. See NOTICE.md.
- Web pages quoted in answers belong to their websites; CAEL links each one.
Limits
- Web answers are only as good as the pages found; CAEL does not check them beyond agreement between websites.
- The facts table is finite, and supported reasoning and programming forms are narrow.
- No picture generation in 0.3.2.