vgi-rpc

TypeScript server library for the vgi-rpc framework. Implements RPC servers that communicate over stdin/stdout using Apache Arrow IPC serialization.

Define RPC methods with Arrow-typed schemas, serve them over stdin/stdout, and interact with them using the Python vgi-rpc CLI or any vgi-rpc client. Unlike JSON-over-HTTP, structured data stays in Arrow columnar format for efficient transfer.

Key features:

• Three method types — unary (request-response), producer (server-streaming), and exchange (bidirectional-streaming)
• Apache Arrow IPC wire format — efficient columnar serialization compatible with the Python vgi-rpc framework
• Schema shorthand — declare schemas with { name: str, count: int } instead of manual Schema/Field construction
• Fluent protocol builder — chain .unary(), .producer(), .exchange() calls to define your service
• Type-safe streaming state — generic <S> parameter threads state types through init and produce/exchange functions
• Runtime introspection — opt-in __describe__ method for dynamic service discovery via the CLI
• Result validation — missing required fields in handler results throw descriptive errors at emit time
• Three client transports — HTTP, subprocess, and raw pipe, all sharing a unified RpcClient interface

Installation

# Bun
bun add @query-farm/vgi-rpc

# npm
npm install @query-farm/vgi-rpc

# Deno
deno add npm:@query-farm/vgi-rpc

Runtime Compatibility

vgi-rpc works with Bun, Node.js 22+, and Deno 2+. The HTTP handler and HTTP client are fully runtime-agnostic. The subprocess transport (subprocessConnect) requires Bun.

Browser Usage

The HTTP client works in browsers. Use httpConnect to call a vgi-rpc server from any browser application:

import { httpConnect } from "@query-farm/vgi-rpc";

const client = httpConnect("https://api.example.com");
const result = await client.call("add", { a: 2, b: 3 });

You will need a bundler (Vite, esbuild, webpack, etc.) that can resolve the @query-farm/apache-arrow dependency. Server-only exports (VgiRpcServer, createHttpHandler, subprocessConnect, pipeConnect) are not available in browsers.

Quick Start

import { Protocol, VgiRpcServer, str, float } from "@query-farm/vgi-rpc";

const protocol = new Protocol("Calculator");

protocol.unary("add", {
  params: { a: float, b: float },
  result: { result: float },
  handler: async ({ a, b }) => ({ result: a + b }),
  doc: "Add two numbers.",
});

protocol.unary("greet", {
  params: { name: str },
  result: { result: str },
  handler: async ({ name }) => ({ result: `Hello, ${name}!` }),
  doc: "Greet someone by name.",
});

const server = new VgiRpcServer(protocol, { enableDescribe: true });
server.run();

Client

Connect to any vgi-rpc server programmatically:

import { httpConnect, subprocessConnect } from "@query-farm/vgi-rpc";

// HTTP transport
const client = httpConnect("http://localhost:8080");
const result = await client.call("add", { a: 2, b: 3 });
console.log(result); // { result: 5 }
client.close();

// Subprocess transport (spawns server, communicates over pipes)
const client2 = subprocessConnect(["bun", "run", "server.ts"]);
const result2 = await client2.call("greet", { name: "World" });
console.log(result2); // { result: "Hello, World!" }
client2.close();

All transports share the same RpcClient interface: call(), stream(), describe(), close().

Testing with the Python CLI

Test it with the Python CLI:

# Describe the service
vgi-rpc --cmd "bun run server.ts" describe

# Call a method
vgi-rpc --cmd "bun run server.ts" call add a=1.0 b=2.0
# {"result": 3.0}

vgi-rpc --cmd "bun run server.ts" call greet name=World
# {"result": "Hello, World!"}

Defining Methods

Unary

Single request, single response:

protocol.unary("add", {
  params: { a: float, b: float },
  result: { result: float },
  handler: async ({ a, b }) => ({ result: a + b }),
  doc: "Add two numbers.",
  defaults: { b: 1.0 },
});

The handler receives parsed parameters and returns a record matching the result schema. The optional defaults field provides default values for omitted parameters.

Producer (Server Streaming)

The server produces multiple output batches. The generic <S> parameter infers state types from the init return value:

protocol.producer<{ limit: number; current: number }>("count", {
  params: { limit: int32, batch_size: int32 },
  outputSchema: { n: int32, n_squared: int32 },
  init: async ({ limit, batch_size }) => ({
    limit,
    current: 0,
    batchSize: batch_size,
  }),
  produce: async (state, out) => {
    if (state.current >= state.limit) {
      out.finish();
      return;
    }
    out.emitRow({ n: state.current, n_squared: state.current ** 2 });
    state.current++;
  },
  doc: "Count from 0 to limit-1.",
  defaults: { batch_size: 1 },
});

produce is called repeatedly. Call out.finish() to end the stream. Mutate state in-place between calls.

Exchange (Bidirectional Streaming)

Client sends batches, server responds one output batch per input batch:

protocol.exchange<{ factor: number }>("scale", {
  params: { factor: float },
  inputSchema: { value: float },
  outputSchema: { value: float },
  init: async ({ factor }) => ({ factor }),
  exchange: async (state, input, out) => {
    const value = input.getChildAt(0)?.get(0) as number;
    out.emitRow({ value: value * state.factor });
  },
  doc: "Scale input values by a factor.",
});

Stream Headers

Producer and exchange methods can send a one-time header before the data stream:

protocol.producer<{ count: number; current: number }>("produce_with_header", {
  params: { count: int },
  outputSchema: { index: int, value: int },
  headerSchema: { total_expected: int, description: str },
  headerInit: (params) => ({
    total_expected: params.count,
    description: `producing ${params.count} batches`,
  }),
  init: ({ count }) => ({ count, current: 0 }),
  produce: (state, out) => {
    if (state.current >= state.count) {
      out.finish();
      return;
    }
    out.emitRow({ index: state.current, value: state.current * 10 });
    state.current++;
  },
});

Schema Shorthand

Declare schemas using convenient type singletons instead of manual Schema/Field construction:

import { str, bytes, int, int32, float, float32, bool } from "@query-farm/vgi-rpc";

// Shorthand
protocol.unary("echo", {
  params: { name: str, count: int, value: float },
  result: { result: str },
  handler: ({ name }) => ({ result: name }),
});

// Equivalent verbose form
import { Schema, Field, Utf8, Int64, Float64 } from "@query-farm/apache-arrow";

protocol.unary("echo", {
  params: new Schema([
    new Field("name", new Utf8(), false),
    new Field("count", new Int64(), false),
    new Field("value", new Float64(), false),
  ]),
  result: new Schema([new Field("result", new Utf8(), false)]),
  handler: ({ name }) => ({ result: name }),
});

Type singletons

| Singleton | Arrow Type | Python equivalent | |-----------|-----------|-------------------| | str | Utf8 | str | | bytes | Binary | bytes | | int | Int64 | int | | int32 | Int32 | Annotated[int, ArrowType(pa.int32())] | | float | Float64 | float | | float32 | Float32 | Annotated[float, ArrowType(pa.float32())] | | bool | Bool | bool |

For complex types (List, Map, Dictionary, nullable fields), use the full Schema/Field constructors from @query-farm/apache-arrow.

Emitting Output

The OutputCollector provides three ways to emit data:

// Column arrays — most efficient for multi-row batches
out.emit({ name: ["alice", "bob"], value: [1.0, 2.0] });

// Single-row convenience
out.emitRow({ name: "alice", value: 1.0 });

// Pre-built RecordBatch
out.emit(batch);

Int64 columns automatically coerce JavaScript Numbers to BigInt.

Client-Directed Logging

Handler functions can emit log messages that travel over the wire to the client:

protocol.unary("process", {
  params: { data: str },
  result: { result: str },
  handler: (params, ctx) => {
    ctx.clientLog("INFO", `Processing: ${params.data}`);
    ctx.clientLog("DEBUG", "Transform complete", { detail: "extra info" });
    return { result: params.data.toUpperCase() };
  },
});

In streaming methods, use out.clientLog():

produce: (state, out) => {
  out.clientLog("INFO", `Producing batch ${state.current}`);
  out.emitRow({ value: state.current });
  state.current++;
},

Error Handling

Exceptions thrown in handlers are propagated to the client as RpcError:

handler: async ({ a, b }) => {
  if (b === 0) throw new Error("Division by zero");
  return { result: a / b };
},

Errors are transmitted as zero-row Arrow batches with EXCEPTION-level metadata. The transport remains clean for subsequent requests.

Testing with the Python CLI

The vgi-rpc CLI can introspect and call methods on any TypeScript server:

pip install vgi-rpc[cli]

# Describe the service
vgi-rpc --cmd "bun run examples/calculator.ts" describe

# Unary call
vgi-rpc --cmd "bun run examples/calculator.ts" call add a=2.0 b=3.0

# Producer stream
vgi-rpc --cmd "bun run examples/streaming.ts" call count limit=5 --format table

# Exchange stream
vgi-rpc --cmd "bun run examples/streaming.ts" call scale factor=2.0

Wire Protocol Compatibility

This library implements the same wire protocol as the Python vgi-rpc framework:

• Multiple sequential Arrow IPC streams on stdin/stdout
• Request batches carry method name and version in batch metadata
• Lockstep streaming: one output batch per input batch
• Zero-row batches for log messages and errors
• __describe__ introspection method for cross-language service discovery

See the Wire Protocol Specification for the full protocol details.

Examples

| Example | Description | |---------|-------------| | calculator.ts | Unary methods: add, multiply, divide | | greeter.ts | String parameters and results | | streaming.ts | Producer and exchange stream patterns | | conformance.ts | 46-method reference service for wire-protocol conformance testing |

Development

# Install dependencies
bun install

# Run tests
bun test

# Build (types + bundle)
bun run build

License

Copyright 2026 Query.Farm LLC https://query.farm. All rights reserved. See LICENSE.md.