Chord AI Technical Overview

this guide explains how chord ai works under the hood, from data connection through to generating answers it's designed for both brand users and technical teams who want to understand what's happening at each step how chord ai works chord ai has two modes classic copilot and copilot next classic copilot follows a structured pipeline to answer questions copilot next uses an agentic model where an llm reasons through your question and calls tools directly both are grounded in the same data and context you've configured classic copilot 1\ connect data source what happens chord securely connects to your commerce data warehouse why it matters chord ai works directly with your data, not a copy of it this ensures answers are always based on your actual store data behind the scenes the connection uses warehouse sharing with validation checks for schema, freshness, and permissions chord ai supports snowflake, bigquery, and redshift note that connecting multiple data sources is not supported at this time 2\ data source modeling what happens your data schema is indexed and transformed into structured, analytics ready models you can enrich this layer with business specific logic via chord ai context by adding instructions and sql pairs to define how key metrics are calculated why it matters this is the abstraction layer between your warehouse and chord's ai models it ensures terms like "revenue" or "customer" mean the same thing every time, across every query behind the scenes chord ai indexes your schema and stores it in a vector database we only work with the schema, not the underlying data itself, keeping your models decoupled from the warehouse 3\ intent interpretation and routing what happens when a question is submitted via copilot chat, chord ai interprets the intent behind it before retrieving any data it determines what the question is really asking, identifies any ambiguity, and routes the request to the appropriate data and logic why it matters this step is what allows copilot to handle complex or loosely worded questions accurately rather than treating every input as a literal string, chord ai reasons about what you're trying to understand before it acts behind the scenes the clarifier agent may activate at this stage to ask follow up questions if the query needs more context once intent is confirmed, the request is passed to the retrieval layer 4\ data retrieval what happens chord ai retrieves the relevant data tables from the modeling layer and pulls live data from your warehouse at query time why it matters real time retrieval means answers reflect your most current data, not a cached snapshot from a prior sync behind the scenes chord ai queries the vector database to identify which schema elements are relevant, pulls the top results, then enriches the query with live warehouse data before passing it to the llm 5\ generating outputs via llm what happens chord ai builds an optimized sql query, validates it, runs it against your data source, and returns the result through copilot why it matters this step automates complex analysis, giving you accurate answers without writing sql manually behind the scenes chord ai uses two anthropic claude models claude opus handles sql generation, where reasoning depth matters most claude haiku handles response generation and classification, optimized for speed and consistency for more detail, see chord ai models powered by anthropic docid\ sse5tr6ceztltsvfgxa6i copilot next copilot next is a ground up rebuild of the copilot experience, currently in early access rather than following a fixed pipeline, copilot next uses an llm agent that reasons through your question and decides which tools to call, making it better suited for complex, multi part questions how it works you ask a question in copilot next an llm agent receives your question along with access to a set of chord ai tools via chord mcp, a model context protocol server that exposes your schema, saved views, sql pairs, instructions, and sql execution as individual callable tools the agent chains these tools as needed searching your schema, retrieving relevant context, and executing sql against your warehouse results stream back in real time as the agent composes its answer classic copilot vs copilot next classic copilot copilot next orchestration fixed pipeline (intent, retrieval, sql, answer) llm agent that chains tools dynamically data access bundled inside the pipeline individual mcp tools via chord mcp best for straightforward data questions complex, multi step, or exploratory questions status generally available early access copilot next is currently available to a limited set of customers to join the waitlist, reach out to product\@chord co mailto\ product\@chord co data privacy chord ai does not store your data we work with your schema only, meaning the structure and definitions of your data, not the records themselves your data never leaves your warehouse without your explicit configuration for more detail on data security, see the chord ai faq docid\ zpwngfj3ck6uwam29tio9