Skip to Content
sdkexamplespythonbeam-searchREADME

Last Updated: 3/9/2026

Beam Search Example

Demonstrates beam search decoding for higher-quality text generation.

Beam search is a decoding strategy that maintains multiple candidate sequences (beams) at each step, selecting the most likely overall sequences rather than greedily choosing the best token at each position.

Setup (One-Time)

From the pie repository root:

cd sdk/python
uv venv
source .venv/bin/activate
uv pip install -e ".[dev]"
uv pip install -e ../tools/bakery

Build

From the pie repository root (with venv activated):

# Build
bakery build "$PWD/sdk/examples/python/beam-search" \
    -o "$PWD/beam-search.wasm"

Run

Requires a running Pie engine:

# Submit with default settings (beam_size=4)
pie-cli submit beam-search.wasm -- --prompt "What is 2 + 2?"
 
# Submit with custom beam size
pie-cli submit beam-search.wasm -- \
    --prompt "What is 2 + 2?" \
    --beam_size 8 \
    --max_tokens 256

Arguments

  • prompt: The input prompt (default: “Explain the LLM decoding process ELI5.”)
  • beam_size: Number of candidate sequences to maintain (default: 4)
  • max_tokens: Maximum tokens to generate (default: 128)
  • system: System prompt (default: “You are a helpful, respectful and honest assistant.”)
READMEREADME