Pipeline API Reference¶

API reference for the pipeline module -- the orchestration layer of anchor.

ContextPipeline¶

The main orchestrator that assembles context from multiple sources into a token-aware, priority-ranked context window.

from anchor import ContextPipeline

class ContextPipeline:
    def __init__(
        self,
        max_tokens: int = 8192,
        tokenizer: Tokenizer | None = None,
        budget: TokenBudget | None = None,
    ) -> None: ...

Parameters:

Raises: ValueError if max_tokens <= 0.

Properties¶

Methods¶

add_step(step) -> ContextPipeline¶

Add a pipeline step. Returns self for chaining.

def add_step(self, step: PipelineStep) -> ContextPipeline: ...

with_memory(memory) -> ContextPipeline¶

Attach a memory provider. Any object satisfying the MemoryProvider protocol (i.e. having a get_context_items() -> list[ContextItem] method) is accepted.

def with_memory(self, memory: MemoryProvider) -> ContextPipeline: ...

with_budget(budget) -> ContextPipeline¶

Attach a TokenBudget for fine-grained allocation.

def with_budget(self, budget: TokenBudget) -> ContextPipeline: ...

with_formatter(formatter) -> ContextPipeline¶

Set the output formatter.

def with_formatter(self, formatter: Formatter) -> ContextPipeline: ...

add_system_prompt(content, priority=10) -> ContextPipeline¶

Add a system prompt as a high-priority context item with source=SourceType.SYSTEM.

def add_system_prompt(self, content: str, priority: int = 10) -> ContextPipeline: ...

add_callback(callback) -> ContextPipeline¶

Register an event callback for pipeline observability.

def add_callback(self, callback: PipelineCallback) -> ContextPipeline: ...

with_query_enricher(enricher) -> ContextPipeline¶

Attach a query enricher for memory-aware query expansion. The enricher is called after memory items are collected but before pipeline steps execute.

def with_query_enricher(self, enricher: ContextQueryEnricher) -> ContextPipeline: ...

step(fn=None, *, name=None, on_error="raise")¶

Decorator to register a synchronous function as a pipeline step. Usable with or without arguments: @pipeline.step or @pipeline.step(name="x", on_error="skip").

Raises: TypeError if the function is async (use async_step instead).

async_step(fn=None, *, name=None, on_error="raise")¶

Decorator to register an async function as a pipeline step. Same usage pattern as step.

Raises: TypeError if the function is not async.

build(query) -> ContextResult¶

Execute the full pipeline synchronously and return assembled context.

def build(self, query: str | QueryBundle) -> ContextResult: ...

Accepts either a plain string (auto-wrapped in QueryBundle) or a QueryBundle.

abuild(query) -> ContextResult¶

Execute the full pipeline asynchronously. Supports both sync and async steps: sync steps are called directly, async steps are awaited.

async def abuild(self, query: str | QueryBundle) -> ContextResult: ...

PipelineStep¶

A single composable step in the context pipeline. This is a dataclass.

from anchor import PipelineStep

@dataclass(slots=True)
class PipelineStep:
    name: str
    fn: SyncStepFn | AsyncStepFn
    is_async: bool = False
    on_error: Literal["raise", "skip"] = "raise"
    metadata: dict[str, Any] = field(default_factory=dict)

Fields:

Methods¶

execute(items, query) -> list[ContextItem]¶

Execute the step synchronously. Raises TypeError if the step is async.

aexecute(items, query) -> list[ContextItem]¶

Execute the step asynchronously. Works for both sync and async step functions.

Factory Functions¶

retriever_step(name, retriever, top_k=10) -> PipelineStep¶

Create a step from a Retriever protocol implementation. Appends retrieved items to the current list.

from anchor import retriever_step
step = retriever_step("search", my_retriever, top_k=5)

async_retriever_step(name, retriever, top_k=10) -> PipelineStep¶

Async variant. Wraps an AsyncRetriever (must have aretrieve(query, top_k)).

from anchor import async_retriever_step
step = async_retriever_step("async-search", my_async_retriever, top_k=5)

filter_step(name, predicate) -> PipelineStep¶

Create a step that filters items by a predicate function. Items where the predicate returns False are removed.

from anchor import filter_step
step = filter_step("score-gate", lambda item: item.score > 0.5)

postprocessor_step(name, processor) -> PipelineStep¶

Create a step from a PostProcessor protocol implementation.

from anchor import postprocessor_step
step = postprocessor_step("dedup", my_deduplicator)

async_postprocessor_step(name, processor) -> PipelineStep¶

Async variant. Wraps an AsyncPostProcessor (must have aprocess(items, query)).

reranker_step(name, reranker, top_k=10) -> PipelineStep¶

Create a step from a Reranker protocol implementation. The reranker scores and returns the top-k most relevant items.

from anchor import reranker_step
step = reranker_step("rerank", my_reranker, top_k=3)

async_reranker_step(name, reranker, top_k=10) -> PipelineStep¶

Async variant. Wraps an AsyncReranker (must have arerank(query, items, top_k)).

query_transform_step(name, transformer, retriever, top_k=10) -> PipelineStep¶

Create a step that expands the query into multiple variants, retrieves for each, and merges results using Reciprocal Rank Fusion (RRF). New items are deduplicated by ID.

from anchor import query_transform_step
step = query_transform_step("multi-query", my_transformer, my_retriever, top_k=5)

classified_retriever_step(name, classifier, retrievers, default=None, top_k=10) -> PipelineStep¶

Create a step that classifies the query and routes to the appropriate retriever.

from anchor import classified_retriever_step
step = classified_retriever_step(
    "router",
    classifier=my_classifier,
    retrievers={"technical": tech_retriever, "general": general_retriever},
    default="general",
    top_k=5,
)

Raises: RetrieverError if the label has no matching retriever and no default.

auto_promotion_step(extractor, store, consolidator=None, name="auto_promotion", on_error="skip") -> PipelineStep¶

Create a step that extracts and stores memories from context. This is a side-effect-only step that returns items unchanged.

graph_retrieval_step(graph, store, entity_extractor, ...) -> PipelineStep¶

Create a step that retrieves memory entries linked to graph entities via BFS traversal.

create_eviction_promoter(extractor, store, consolidator=None) -> Callable¶

Create an on_evict callback that promotes evicted conversation turns to long-term memory. Designed for SlidingWindowMemory(on_evict=...).

from anchor import create_eviction_promoter, SlidingWindowMemory

promoter = create_eviction_promoter(extractor, store, consolidator)
memory = SlidingWindowMemory(max_tokens=4096, on_evict=promoter)

Returns: A callable with signature (list[ConversationTurn]) -> None.

PipelineCallback¶

A runtime-checkable protocol for pipeline event callbacks. All methods are optional -- implement only the ones you need.

from anchor import PipelineCallback

class PipelineCallback(Protocol):
    def on_pipeline_start(self, query: QueryBundle) -> None: ...
    def on_step_start(self, step_name: str, items: list[ContextItem]) -> None: ...
    def on_step_end(self, step_name: str, items: list[ContextItem], time_ms: float) -> None: ...
    def on_step_error(self, step_name: str, error: Exception) -> None: ...
    def on_pipeline_end(self, result: ContextResult) -> None: ...

See Also¶

• Pipeline Guide -- walkthrough with complete examples
• Models API Reference -- ContextItem, ContextResult, TokenBudget
• Exceptions Reference -- error classes