josephwoodall/uchi

Uchi is a Python library that learns from text and answers questions without using a neural network or a language model. Instead of the statistical weights that most AI systems rely on, it stores information in a structure called a prefix trie, which is closer to an indexed lookup table than to a trained model. When you ask it a question, it searches that structure and returns the most likely next words based on what it has seen before. The practical result of this design is that Uchi never forgets facts it has been taught. Because new information is added to the trie without touching existing entries, loading a thousand new facts does not erase the ones already stored. Benchmarks in the repository show that ten anchor facts remain perfectly recallable after one thousand unrelated facts are streamed in on top. Response time stays roughly constant as the knowledge base grows, because the lookup time depends on the depth of the query in the trie, not on the total number of facts stored. You interact with Uchi in three ways: a terminal chat interface you launch by typing uchi in your shell, a REST API you can run locally and call with standard HTTP requests, or a Python library you import directly into your own code. All three share the same brain file on disk, so anything you teach through one interface is available through the others. You can also run bootstrap scripts to pre-load Wikipedia facts, Python standard library patterns, or knowledge from public datasets before distributing your brain file to others. The repository describes Uchi as a deterministic sequence predictor. It claims zero hallucination because it only produces output from sequences it has actually seen. The trade-off is limited generalization: code completion scores are low on standard benchmarks because the system has not been pre-trained on large code corpora. The README is transparent about this, presenting the benchmarks alongside their limitations. Uchi is available as a Python package via pip and is MIT licensed. The full README is longer than what was shown.