Fabrile

The product

Fabrile adds a chat agent to a website. A customer points it at their help docs, website, Zendesk, Notion or Google Drive, and the agent answers visitors from that content, citing the source. It goes live with one line of embed code or through the Shopify and Squarespace apps.

What runs on Ingram Cloud

Fabrile has no model integration of its own. Its code covers the agent editor, the content connectors, the chat widget and billing. Model calls to GPT, Claude and Gemini, conversation history and knowledge-base search run on Ingram Cloud.

Agents and versions

Each agent a customer builds in Fabrile is an Ingram Cloud agent, matched by a stable slug. Publishing in Fabrile's editor publishes a new agent version and rolls it out to every smith running that agent.

A smith per visitor

Each website visitor gets their own smith of the agent, with threads and memory no other visitor can reach. History is stored server-side, so each turn sends only the new message. With memory on, the agent remembers a returning visitor.

Knowledge bases are vector stores

Each Fabrile knowledge base is an Ingram Cloud vector store attached to the agent. Crawled pages and uploaded documents are ingested there, and retrieval runs inside the run through file_search, so the sources a visitor sees come from the step that produced the answer. Fabrile runs no embedding pipeline or vector database of its own.

The client

A chat turn is a Vercel AI SDK call through @ingram-cloud/ai-sdk on the OpenAI-compatible Responses surface, with tools for Fabrile's connectors declared on the request. Conversation titles and suggested agent instructions are generated on Ingram Cloud too.