PhiNetwork Chatbot

A small Rust experiment around a PhiNetwork-style learner and a command-line chatbot.

The model learns response classes from text examples. It tokenizes input words, builds phi terms from active features, and trains either dense learned curves or sparse phi components. The chatbot can remember new examples between sessions and can draw the learned overall phi curve as ASCII.

Run

cargo run

At startup, choose a model:

1. dense curve   - smoother, original curve PhiNetwork, poor scaling
2. sparse scalar - scalable, more exact lexical matching
3. sparse curve  - scalable sparse terms with learned curves

Press Enter to use the default sparse curve mode.

CLI Commands

add <message> => <reply>     add one training example and save it
train [epochs] [epsilon]     rebuild and train the chatbot
ask <message>                ask the trained chatbot
suggest [limit] <message>    list likely replies from remembered examples
examples                     list training examples
responses                    list learned response classes
clear context                forget accumulated session phi terms
curve                        draw the learned overall phi curve
keypair [shares]             print encoded phi and encrypted phin shares
tokens <message>             show word tokens for a message
vocab                        list bag-of-words features
help                         show command help
quit                         exit

Plain text without a command is treated like ask <message>.

If the chatbot is not confident enough, it asks for the right response and remembers that answer. Run train to rebuild the learned phi state from remembered examples. When an answer is confident, recursion feeds that answer back as the next prompt and prints each confident memorized transition. It stops when the next answer is low confidence, repeats, lacks an exact remembered transition from the previous answer, or reaches the recursion cap.

Example Session

> add rust ownership borrowing => Rust ownership controls memory without a garbage collector.
added and remembered example. Run `train` to update the model.
> train
trained 22 examples into 10 responses with 34 word features
> ask explain rust borrowing
Rust ownership controls memory without a garbage collector. (0.742)
> curve
phi_all
    ...

Persistence

The chatbot stores learned data under data/:

data/chatbot_memory.tsv       remembered training examples
data/chatbot_phi_all.tsv      learned phi state

Memory examples are loaded on startup. Sparse phi state is also loaded on startup when compatible with the selected mode. The phi state also stores an encoded phi_all export line for inspection.

The keypair [shares] command is inspired by BLS12-381 key material and curve points, but it is an application-specific encoding of phi_all points. It prints the direct encoded phi_all points, then n encrypted phin shares whose component formulas show how each share is composed, for example phi(a) + phi(ab) + phi(ac) when n = 3. The encrypted phin shares are masked additive shares: all n unmasked shares combine to recover encoded phi_all. It should not be treated as a standard BLS signature key, wallet key, or general-purpose identity secret.

Project Layout

src/main.rs                    entrypoint
src/chatbot.rs                 chatbot/session/persistence orchestration
src/phinetwork.rs              dense PhiNetwork core
src/commands/                  one module per CLI command
src/classifiers/               classifier implementations
src/classifiers/dense_curve.rs dense curve classifier
src/classifiers/sparse_phi.rs  sparse scalar and sparse curve classifier
src/classifiers/curve_plot.rs  ASCII curve plotting helpers

Development

cargo fmt --check
cargo test

License

MIT. See LICENSE.