slot-flight

Slot-wise LLM value streaming with server-owned JSON assembly.

slot-flight does not ask the model to stream a valid JSON document. The model streams compact slot frames, and an SDK maps those frame ids to JSON paths, validates each value, retries failed slots, and assembles the final object itself.

<1>
Alice
</1>
<2>
Senior Engineer
</2>
<3:0>
streaming
</3:0>
<3:1>
json assembly
</3:1>

Structural Streaming

Nested objects and arrays stream as individual raw slot values. Array fields use indexed frame tags such as <3:0>...</3:0>, and arrays of objects stream each item field separately.

This lets the SDK assemble the final object while preserving typed validation and partial updates during streaming.

Dynamic key/value objects such as records, maps, and mappings cannot be inferred as structural slots because their keys are not known ahead of streaming. Model them as explicit fields or handle them outside slotObject() / slot_object().

Repository Layout

This repository is organized as a multi-language SDK workspace:

• docs/protocol: language-neutral slot frame protocol
• packages/typescript: TypeScript SDK and provider adapters
• packages/python: Python SDK core

Each SDK should keep the same protocol behavior while using language-native schema and provider integration patterns.

TypeScript

bun add slot-flight zod

Provider SDKs stay in your application:

bun add openai
# or
bun add ai @ai-sdk/openai

Define the output shape with Zod. Every generated field is registered through .describe(), which becomes the model-facing slot instruction.

import OpenAI from "openai";
import { z } from "zod";
import { slotObject } from "slot-flight";
import { streamSlotObject } from "slot-flight/adapters/openai";

const openai = new OpenAI({
  apiKey: process.env.API_KEY,
  baseURL: "https://integrate.api.nvidia.com/v1"
});

const stream = streamSlotObject({
  client: openai,
  model: "openai/gpt-oss-20b",
  messages: [
    {
      role: "user",
      content: "Classify this customer support feedback for a triage queue."
    }
  ],
  temperature: 0.2,
  output: slotObject({
    schema: z.object({
      summary: z
        .string()
        .min(1)
        .describe("Write one concise operational summary."),
      sentiment: z
        .enum(["positive", "neutral", "negative", "mixed"])
        .describe("Write exactly one of: positive, neutral, negative, mixed."),
      priority: z
        .enum(["low", "medium", "high"])
        .describe("Write exactly one of: low, medium, high.")
    })
  })
});

for await (const slot of stream.completedSlotStream) {
  console.log(slot.slot, slot.value);
}

const finalObject = await stream.finalObject;

TypeScript-specific docs and examples live in the TypeScript SDK guide and packages/typescript.

Python

uv add slot-flight

Optional adapter dependencies:

uv add "slot-flight[openai]"
uv add "slot-flight[openai-compatible]"
uv add "slot-flight[langchain]"

The Python SDK provides a Pydantic-first object API plus provider/framework adapters for the OpenAI SDK, OpenAI-compatible HTTP endpoints, and LangChain.

from pydantic import BaseModel, Field
from slot_flight import slot_object
from slot_flight.adapters.openai import stream_slot_object


class Triage(BaseModel):
    summary: str = Field(description="Write one concise operational summary.")
    tags: list[str] = Field(description="Write exactly 3 tags, one per frame.")


stream = stream_slot_object(
    client=openai,
    model="gpt-4.1-mini",
    messages=[{"role": "user", "content": "Classify this feedback."}],
    output=slot_object(Triage),
)

async for slot in stream.completed_slot_stream():
    print(slot.slot, slot.value)

result = await stream.final_object()

Python package details and provider examples live in the Python SDK guide and packages/python.

Protocol

The shared protocol is documented in docs/protocol/slot-frame-protocol.md.

Contributing

Contribution, local verification, CI, and release notes are documented in CONTRIBUTING.md.