C3 AI Documentation Home

Configure the Embedder

The Embedder generates a vector embedding for each document chunk and writes it to the vector store. The Embedder is required. Without it, the pipeline can't complete and downstream search has nothing to query.

Open the Embedder configuration

On the canvas, select the Embedder node. The Configure Embedder panel opens.

Name the Embedder

At the top of the panel, the Name field is prepopulated with <source-collection-name> Embedder. Keep the default or enter a different name. The canvas node title shows this name.

Select an embedding model

The Embedding Model dropdown lists every embedding client configured in your environment. The composition of the dropdown depends on which clients are registered in Model Hub > Model Clients. Common options include Multilingual E5 Large, one or more vendor-provided clients, and a default (CONFIGURE ME) placeholder.

Use Multilingual E5 Large when:

  • You're indexing a multilingual document set.
  • You want a balanced model that handles short queries and long passages.
  • You're prototyping and plan to tune later.

Use a vendor-specific embedding client (such as a Cohere, Amazon Titan, or OpenAI embedding model) when:

  • Your downstream retrieval system, such as GenAI Search or a custom RAG service, is built around that model's vector space.
  • You have an organization standard for embeddings.
  • You've benchmarked a vendor model against your evaluation set.

Use the default (CONFIGURE ME) placeholder when:

  • Your platform administrator has registered an embedder placeholder that points to a model you configure later.
  • You're stubbing out a pipeline that you'll wire to a real model before production.

Don't leave the default (CONFIGURE ME) placeholder in a production pipeline.

The exact set of available models varies between environments. The dropdown lists what's available locally, not a universal catalog.

When to include metadata in embeddings

The Include Metadata in Embeddings section — with its Document Tags toggle — controls whether the platform concatenates the chunk text with metadata tags before embedding. The section description reads Enhances semantic matching with metadata context.

The Document Tags toggle is gated on having a Metadata Tagger configured in the pipeline. When the pipeline has no Metadata Tagger, the toggle is disabled and an amber banner reads Configure Metadata Tagger to enable Document Tags. Configure the Metadata Tagger first, then return to the Embedder to switch the toggle on.

Switch the toggle on when:

  • Your pipeline runs the Metadata Tagger and you want tag values to influence semantic search results.
  • You're tagging documents with attributes such as topic, department, or document type that describe the chunk in a meaningful way.

Leave the toggle off when:

  • You want filtering on metadata to happen at query time, not at embedding time.
  • You're concerned that tag values might overpower the document text in the vector space.

Name the Retriever ID

Each Embedder writes to a vector store identified by its Retriever ID. The default is default-pg.

Use the Retriever ID to:

  • Reference this UDI pipeline from a downstream GenAI Search retriever or RAG agent.
  • Separate vector stores for each use case so that an employee handbook retriever and a product manual retriever don't share results.
  • Version pipelines, such as policies_v1 or policies_v2, when you want to re-index without disrupting existing queries.

Recommended naming conventions:

  • Use lowercase words separated by underscores. For example, quarterly_reports_2026.
  • Include the corpus and version, not the model. The model can change.
  • Keep the name short enough to type in downstream configuration.

Save the configuration

Select Save. The canvas creates the vector-store entity node downstream of the Embedder automatically. The pipeline now has all the required nodes.

Re-index after Embedder changes

If you change the embedding model, the Retriever ID, or the metadata configuration after the pipeline has already run, the platform marks every previously-indexed file as Needs Re-indexing. Re-execute the pipeline to refresh the vector store with the new embeddings.

See also

Was this page helpful?