Python Function for a Basic Gen Model

JPMorganChase is evaluating lightweight text-generation components for internal AI assistants. Implement a token-level generative model without external machine learning libraries.

Given a training corpus, build an interpolated n-gram model that predicts the next token from the longest available context. Apply additive smoothing so every vocabulary token remains possible, back off to shorter contexts when a context was not observed, and generate a reproducible continuation using a seeded random number generator.

Formal Specification

Implement generate_text(corpus, prompt, order, max_tokens, temperature, seed).

• corpus and prompt are non-empty whitespace-separated strings.
• order is the maximum context length.
• max_tokens is the number of tokens to generate, excluding the prompt.
• temperature controls randomness. A value of 0 selects the highest-probability token; positive values sample from temperature-adjusted probabilities.
• seed initializes Python's pseudorandom generator.
• Return a string containing exactly the generated tokens separated by single spaces. Return an empty string when max_tokens is zero.

Use additive smoothing with a fixed alpha of 0.1. If several tokens have the same highest probability at temperature 0, choose the lexicographically smallest token.