Menu

Mattering Automations Voyage AI integration

Connect Voyage AI to Mattering Automations: text embeddings, document reranking, multimodal embeddings, and contextualized chunk embeddings as workflow steps, authenticated with a Voyage API key.

The Mattering Automations Voyage AI integration adds Voyage's retrieval models to your automation canvas: text embeddings, document reranking, multimodal embeddings over text mixed with image or video URLs, and contextualized chunk embeddings that keep RAG chunks aware of their surrounding document.

Setup takes a couple of minutes: create an API key in the Voyage dashboard, attach it as a credential in Mattering, and test the connection. This guide walks through each step and ends with the node reference.

The integration is four nodes because that is Voyage's entire model API, and all four are pure compute: they read inputs and return vectors or rankings, storing nothing on Voyage's side. The only thing to manage is token spend.

Before you start

  • A Voyage AI account and API key from the dashboard (dashboard.voyageai.com).
  • Token balance or billing set up. New accounts include free trial tokens; after that, usage bills per million tokens by model.
  • A Mattering workspace with access to the Automations app.

Create a Voyage AI API key

Mattering authenticates to Voyage with an API key, sent as a bearer credential on every request. There are no scopes; a key covers all four endpoints.

Sign in at dashboard.voyageai.com and create an API key.
Copy the key.
Voyage has no free read endpoint, so any authentication check costs something. The integration keeps that honest and tiny: the connection test embeds one short word on Voyage's lightest model, a real end-to-end check whose cost rounds to nothing.

Connect Voyage AI to Mattering

In Mattering Automations, open Connectors and click Add Connector.
Choose Voyage AI from the list.
Click Create new credential, paste your API key, and save.
The credential is stored encrypted in Google Secret Manager and never appears in workflow logs.
Click Test Connection: Mattering embeds a single word on Voyage's lightest model and confirms the key authenticates.
Click Connect. All 4 Voyage AI nodes appear in your canvas palette.

Verify it works

After connecting, open any automation canvas and check the palette: a Voyage AI section should list the four nodes. Drop Voyage AI: Create Text Embeddings onto the canvas, give it a short sentence and a model name, and run a test step: a successful response returns a vector and reports token usage.

To see the retrieval side, run Voyage AI: Rerank Documents with a question and three or four short candidate texts: the response returns them reordered with relevance scores, which is the shape most workflows build on.

Node reference

4 nodes, generated from the connector itself so this reference always matches what ships.

Action nodes 4

Voyage AI: Create Contextualized Chunk Embeddings

voyage_create_contextualized_embedding

Embed all the chunks of a document together so each chunk's vector carries its surrounding document context. Fixes the RAG blind spot where isolated chunk embeddings lose what their references mean.

Voyage AI: Create Multimodal Embeddings

voyage_create_multimodal_embedding

Embed inputs that interleave text with image or video URLs into one vector space, so a text query can retrieve a diagram or product photo. Media referenced by URL; raw bytes never pass through the workflow.

Voyage AI: Create Text Embeddings

voyage_create_embedding

Turn text into vectors, with input type hints (query versus document), output dimension and quantization options. The connector's auth probe: the connection test embeds one short word on Voyage's lightest model.

Voyage AI: Rerank Documents

voyage_rerank

Order candidate documents by relevance to a query, with scores. The quality step between a wide vector-search shortlist and the documents an LLM or person actually sees.

Using the integration

On the automation canvas

Voyage nodes chain with every other node: records in a Mattering database can carry embeddings refreshed when content changes, and Voyage AI: Rerank Documents can decide which knowledge article, contract clause, or ticket a workflow surfaces to a person, with a score threshold routing weak matches to a human instead.

From AI assistants over MCP

Every node speaks MCP, the open protocol for connecting AI assistants to tools. Claude, ChatGPT, and Matter can embed queries and rerank documents through the same nodes your workflows use. All four nodes are pure compute, so the whole surface is safe to hand an agent; the only budget is tokens.

Query versus document embeddings

Voyage AI: Create Text Embeddings takes an input type hint that tells Voyage whether it is embedding a search query or a document. Voyage's retrieval models embed the two sides differently for better matching, so set it: query at question time, document at indexing time. Output dimension and quantization options are there for teams tuning storage against accuracy.

Reranking as the quality step

The standard pattern: a cheap vector search produces a wide shortlist, Voyage AI: Rerank Documents tightens it with real relevance scores, and only the top documents reach an LLM or a person. It is the highest-leverage single step in most retrieval stacks, and on the canvas it is one node between the search and the answer.

Multimodal and contextualized embeddings

Voyage AI: Create Multimodal Embeddings embeds inputs interleaving text with image or video URLs into one vector space, so a text query can retrieve a diagram or product photo; media is referenced by URL and raw bytes never pass through the workflow. Voyage AI: Create Contextualized Chunk Embeddings embeds a document's chunks together so each vector carries surrounding context, which matters most for contracts, policies, and documentation where chunks refer to things defined elsewhere.

Troubleshooting

Test Connection fails with a 401
The key is invalid, was revoked, or picked up stray whitespace when pasted. Create a fresh key in the Voyage dashboard and update the credential.
Test Connection fails but the key is definitely valid
Because Voyage has no free endpoint, the connection test performs a real one-word embedding, so it can also fail for non-auth reasons: an exhausted token balance, a billing hold on the account, or a rate limit. Read the error detail on the test result; a 401 is the key, anything about quota or billing is the account.
Embedding calls fail with a model error
The model field must be a current Voyage model name, and models retire over time. Check the Voyage docs for the current lineup and update the node's model field. Because the model is plain config, no connector update is needed to adopt a new one.
Retrieval quality is worse than expected
The usual causes, in order: the input type hint is not set (query at question time, document at indexing time, and mixing them degrades matching); query and documents were embedded with different models, which produces incompatible vectors; or chunks lost their context, which Voyage AI: Create Contextualized Chunk Embeddings exists to fix. Re-embedding the corpus consistently resolves the first two.
Multimodal embeddings fail on image or video inputs
Media is fetched by Voyage from the URLs you pass, so each URL must be publicly reachable: no logins, no expired signed links, and a direct media response. Host assets somewhere public or refresh the signed URL at workflow time.
Calls fail with a 429 rate limit
Voyage enforces per-model request and token rate limits that scale with account tier. Batch texts into fewer, larger embedding calls rather than one call per text, space out scheduled sweeps, and let the workflow wait and retry.

Voyage AI is a trademark of its respective owner. Mattering uses platform names to identify the services these integrations connect to; no endorsement or affiliation is implied.

Start Free Explore our Products