ambi-pipelines: optional execution strategies

ambi-pipelines keeps higher-level reasoning patterns outside Ambi's small core. Its repository source version is 0.1.1 alongside Ambi 0.4.1. These pipelines implement the Pipeline contract; they are not generic graph nodes and do not automatically compose with every other pipeline.

[dependencies]
ambi = "0.4.1"
ambi-pipelines = "0.1.1"

Use matching workspace path dependencies if the versions are not available from your registry.

Included strategies

Pipeline	How it executes	Key limitation
StandardRagPipeline	Retrieves and packs documents, writes dynamic context, delegates to an inner ChatRunner	Mutates session dynamic context
SelfConsistencyPipeline	Runs branches and selects an aggregated answer	Streaming not implemented
BfsBeamSearchPipeline	Expands/evaluates candidate thought branches	Requires user-provided expander and evaluator
ReflexionPipeline	Runs an actor/evaluator loop, stores critiques using a KV provider	Streaming not implemented

Example: RAG pipeline

use ambi::agent::pipeline::Pipeline;
use ambi_pipelines::rag::StandardRagPipeline;

// my_retriever implements ambi_pipelines::rag::Retriever.
let pipeline = StandardRagPipeline::create(my_retriever)
    .max_context_tokens(4096)
    .concurrency(3);

let answer = pipeline.execute(&agent, &state, input_parts).await?;

The Retriever and Document abstractions belong to ambi-pipelines, not core ambi. RAG packs the retrieved text into AgentState.dynamic_context before using its internal ChatRunner, so the application should account for context freshness and shared-state concurrency.

Why these are separate crates

Ambi provides the execution substrate; an application selects the reasoning strategy. The built-in examples should not impose a universal cognitive architecture on users who only need a small ReAct loop or a custom pipeline.

Leaf availability: The 0.4.1 Leaf feature is mounted on ChatRunner explicitly; these auxiliary pipelines may create internal ChatRunner::default() instances, which remain Leaf-free even if the Cargo feature is compiled. See Leaf and Custom Pipeline.