Skip to main content

Introduction

ObjectWeaver is a network orchestration tool that sits between your application and LLM providers. You describe a pipeline of tasks as a JSON schema: each field is a task with its own model, instructions, and dependencies, and ObjectWeaver routes, validates, and assembles the results into a single clean JSON response.

What ObjectWeaver Does​

You send a JSON schema describing the tasks you need completed. ObjectWeaver treats each field as a node in a task network: routing it to the right model, validating the output, and wiring results together. Complex multi-step workflows become a single schema definition.

Why Schema-Driven Orchestration​

ObjectWeaver orchestrates multiple, focused LLM calls — one per task in your schema. That means more input tokens compared to a single monolithic prompt. That's the tradeoff.

But it's a tradeoff worth making. When each task has a narrow, well-scoped instruction instead of a sprawling multi-part prompt, output quality improves significantly. Each node gets the model's full attention with only the context it needs — no cross-contamination, no drift.

And as inference costs continue to drop, the remaining bottleneck is coordination speed. ObjectWeaver is built for this: by treating independent tasks as parallel nodes in the network, it produces complete, validated results as fast as the underlying models allow. If you need sequential control instead, that's supported too — see Processing Order.

How It Works​

  1. Define: Write a JSON schema where each field describes a task — with an instruction, model, and optional dependencies.
  2. Orchestrate: ObjectWeaver builds a task network from your schema, identifying which nodes can run in parallel and which must wait.
  3. Route: Each task is dispatched to the model you specify (e.g., gpt-4o-mini for simple fields, gpt-4o for complex reasoning).
  4. Validate: Every task output is validated against its type definition.
  5. Assemble: All results are wired together and returned as a single, validated JSON object.

Key Features​

  • Task Network from a Schema — Express an entire multi-step workflow as a single JSON schema.
  • Field-Level Model Routing — Use cheaper models for simple tasks, powerful models where it matters.
  • Parallel Execution — Independent tasks run concurrently to reduce end-to-end latency.
  • Guaranteed JSON — Every node output is validated; no parsing failures or malformed responses.
  • Context Isolation — Each task only sees the context it needs, keeping prompts focused and outputs clean.

Background​

ObjectWeaver started during a dissertation project at the University of Bath in 2024, where reliably connecting multiple LLM calls into a coherent result was a constant headache. It was built to solve that — and has since grown into a full network orchestration layer with model routing, parallel execution, and composable schemas.

Next Steps​