Semantic search is now live in the Airbyte context layer.
This is meant for teams who use Airbyte to sync sources with unstructured data like Gong, Salesforce, Google Drive, Jira, Confluence, or Gmail into Snowflake or BigQuery. With semantic search enabled in the context layer, an agent connected to the Airbyte MCP can search that text by meaning, then use SQL to join, filter, and count the results.
Turn it on for a source, and Airbyte indexes the data you already sync, with the associated credentials.
Why this matters Historically, structured data was the primary feed for enterprise analytics. Airbyte got its start moving vast amounts of these structured records for organizations. Obviously, structured data is incredibly valuable for enterprise agents, and always will be.
But a huge share of organizational knowledge doesn't live in structured records. With traditional extraction methods and pipelines, this knowledge remains effectively invisible to agents. While SQL keyword search is excellent for structured data, unstructured data introduces more ambiguity.
Call transcripts and issue threads are messy and unpredictable. The way humans express issues don’t fall into neat keywords or business terminology.
A customer worried about cost might say, "this feels steep for what we're getting", or "I'd have to fight for budget on this." A user reporting a performance bug might write "the dashboard takes forever after the update" without ever typing "latency."
Pure keyword search isn’t enough to understand the meaning behind words, which means it will miss the relevant business context and struggle to answer the question.
Semantic search overcomes this limitation. Airbyte embeds your question and compares it against the meaning of the transcript or issue text, then returns the most relevant passages, ranked by how closely they relate to your question, with the surrounding context attached so your agent knows who said what.
That’s why we built semantic search into Airbyte’s hybrid search inside the context layer.
How it works Indexing starts from data you've already replicated:
Split. Airbyte reads the text fields marked for search on each source, such as Gong call transcripts or Jira issue descriptions, and breaks them into passages.Embed. Each passage becomes a vector that represents its meaning. Airbyte runs the embedding model on its own infrastructure, so your text isn't sent to a third-party embedding API. The model is multilingual, so passages written in languages other than English are searchable too.Rank. When an agent searches, Airbyte embeds the prompt the same way, retrieves the closest passages, and runs them through a re-ranker before returning them.Search runs in hybrid mode by default, combining semantic and keyword matching. A prompt about billing complaints finds "I got charged twice," and a search for a specific order number or product name still finds the exact match. Each result comes back with the matched passage, its surrounding context, a similarity score, and the record it came from, so the agent can join it to the rest of your warehouse with SQL.
What your team can ask Your team can access the context layer and semantic search via the Airbyte MCP inside Claude, ChatGPT, or any other MCP compatible client. The agent decides when to search, when to run SQL, and when to read live from a source:
"Refunds spiked last week. Search that week's Zendesk Chat conversations for the reasons customers gave, then group the matches by plan using our accounts table." "Which streams in our warehouse capture why customers cancel? Check support conversations, Salesforce notes, and Gong transcripts." "RevOps wants the accounts that raised data residency concerns on calls this quarter. Pull them from Gong and add each account's opportunity stage from Salesforce." "Find the Confluence pages where we defined an active customer, and tell me which definition is most recent." The first request shows how the pieces fit. The agent confirms the spike and its date range with SQL, runs semantic search over the support conversations in that window, and gets back passages tied to specific conversation records. It then joins those records to your accounts table to group them by plan. The analyst gets an answer grounded in what customers wrote, without building a text classifier or a vector store first.
Enable semantic search Here’s how to set up semantic search for a supported data source.
First, your organization admin has to open Context layer in Airbyte Cloud and turn on semantic search. After your org has access, make sure the source you want to search syncs to a Snowflake or BigQuery destination. On the Semantic search page, turn on semantic search for the connection. Airbyte optionally backfills the index from your destination. This may incur compute costs from your destination. Connect your agent to the Airbyte MCP at https://mcp.airbyte.com/mcp and ask a question. Your agent can check which streams are indexed and how far the backfill has gotten. If search isn't on for a source yet, the MCP says so and links you to the page where you turn it on.
Semantic search works on a growing set of sources today, including Gong, Google Drive, Salesforce, Jira, Confluence, Gmail, Outreach, Pylon, and Zendesk Chat.
Embeddings and search indexes live on Airbyte's cloud infrastructure, so that means semantic search is only available in Airbyte Cloud or on Airbyte-managed data planes.
Get started Turn on semantic search for one connection you already sync, connect your agent to https://mcp.airbyte.com/mcp, and ask it the question your SQL couldn't answer. The context layer docs cover setup.
New to Airbyte? Start a 14-day free trial of Airbyte Cloud , which lets you access the context layer and semantic search.