Skip to content

README.md

Latest commit

 

History

History
66 lines (47 loc) · 2.36 KB

README.md

File metadata and controls

66 lines (47 loc) · 2.36 KB
Raw

Moss + Vercel AI SDK Cookbook

A minimal example showing how to plug Moss semantic search into an agent built with the Vercel AI SDK, using the official @moss-tools/vercel-sdk tool wrappers.

What this shows

  • Creating a MossClient and wrapping one of your indexes with mossSearchTool
  • Handing that tool to streamText (token-by-token streaming, the default) or generateText (single-shot answer), so the model can retrieve context before replying
  • A script to seed a small sample index so the example runs end-to-end with no existing Moss data required
  • A small CLI script you can run with your own question as an argument

Prerequisites

  • Node.js 18+
  • A Moss project (project id + project key). You do not need an existing index; the seed script below creates one for you.
  • An OpenAI API key

Install

cd examples/cookbook/vercel-ai-sdk
npm install

Configure

Copy the example env file and fill in your own values:

cp .env.example .env
Variable Description
MOSS_PROJECT_ID Your Moss project id
MOSS_PROJECT_KEY Your Moss project key
MOSS_INDEX_NAME The index the search tool should query
OPENAI_API_KEY Your OpenAI API key
OPENAI_MODEL Optional, defaults to gpt-4o-mini
GENERATE Optional, set to true to use generateText instead of the default streamText

Seed a sample index

If you don't already have a Moss index to query, create one from the sample support-doc snippets in seed_data.ts:

npm run seed

This deletes any existing index named by MOSS_INDEX_NAME in your .env and recreates it.

Run

npm start -- "What is the refund policy?"

By default this uses streamText and prints the answer incrementally as it's generated. Set GENERATE=true in your .env to switch to a single-shot generateText call instead.

How it works

moss_vercel.ts prebinds the search tool to a single index name with mossSearchTool({ client, indexName }), so the model only ever has to supply a query (and optional topK) rather than choosing an index itself. The system prompt instructs the model to always search before answering and to cite what it finds, which keeps responses grounded in your own data instead of the model's general knowledge.

Learn more