@contextrie/core

Dynamic context curation for long-running agent work, packaged as Contextrie's core TypeScript library.

This package is intentionally small. It gives you the building blocks for:

• defining indexed sources
• generating source metadata
• judging source relevance for a task
• composing a tight context from judged sources

It does not handle IO, parsing, crawling, or storage.

Install

npm install @contextrie/core ai zod

pnpm add @contextrie/core ai zod

bun add @contextrie/core ai zod

Runtime model

@contextrie/core is published as ESM TypeScript source.

Use it with:

• Bun
• a TypeScript-aware bundler/runtime
• applications already built around the Vercel AI SDK

If you need plain precompiled JavaScript output for direct Node consumption, build that in your app layer or add a package build step in this repo.

What the package includes

• IndexedSourceBase: base contract for retrieval-ready sources
• DocumentSource, ListSource, ReferenceDocumentSource, ReferenceListSource: source implementations
• IndexingAgent: generates metadata from source content
• JudgeAgent: scores source relevance for a task
• ComposerAgent: condenses judged sources into a single context paragraph

Quick example

import { openai } from "@ai-sdk/openai";
import {
  ComposerAgent,
  DocumentSource,
  IndexingAgent,
  JudgeAgent,
} from "@contextrie/core";

const model = openai("gpt-4.1-mini");

const sources = [
  new DocumentSource(
    "architecture",
    undefined,
    "The indexing pipeline adds metadata to a source while preserving the underlying content for deep judgment.",
  ),
  new DocumentSource(
    "roadmap",
    undefined,
    "Q2 focuses on packaging, npm release hygiene, and developer onboarding.",
  ),
];

const indexing = new IndexingAgent(model);
const indexed = await indexing.add(sources).run();

const judge = new JudgeAgent(model);
const judgments = await judge.from(indexed).run({
  objective: "response",
  input: "What matters for the first npm release?",
});

const judgedSources = Object.fromEntries(
  indexed.map((source) => [
    source.id,
    {
      source,
      decision: judgments[source.id],
    },
  ]),
);

const composer = new ComposerAgent(model);
const context = await composer.from(judgedSources).run({
  objective: "response",
  input: "What matters for the first npm release?",
});

console.log(context);

Package boundaries

• no filesystem or network IO abstractions
• no parsing or source discovery
• no storage or persistence layer
• no framework-specific adapters

API overview

Sources

Sources inherit from IndexedSourceBase and must provide:

• kind
• isIterable
• getContent()
• buildIndexInput()
• buildDeepJudgeInput()

Metadata is optional at construction time. A source without metadata is treated as a draft source.

Indexing

IndexingAgent turns draft sources into indexed sources by generating:

• title
• description
• keypoints

Indexing mutates the same source object by filling in source.metadata.

Judgment

JudgeAgent scores sources against a task and returns:

• score: a number from 0 to 1
• reason: a short factual justification

It supports both:

• shallow judgment from source metadata
• deep judgment from full source content

Composition

ComposerAgent takes judged sources and returns one dense markdown paragraph suited for downstream context injection.

Development

See CONTRIBUTING.md.