Use Vectors in Bots & Workflows
Once a vector is trained, two blocks can search it — one in the bot editor, one in the workflow editor. Both work the same way: give them a query, get back the most relevant chunks of your content.
In a bot: the Vector Store block
Found in the bot editor's Database section.
Add and configure the block
| Setting | What it does |
|---|---|
| Vector Store Name | Pick one of your workspace's vectors from the dropdown |
| Query | The search text — usually a variable holding the user's question, e.g. {{user_question}} |
| Number of Documents | How many matching chunks to return (default 2) |
Save the answer
Open Save answer and map a value to a variable:
| Value to extract | What you get |
|---|---|
| Chunk Text | All matching chunks as one text, each prefixed with its metadata (chunk id, section, source URL) — ready to paste into an AI prompt |
| Full Response | A structured list with chunk_text, section, source_url, score, and more — for advanced use |
Feed it to an AI block
The classic pattern:
- Text input block → saves the question in
{{user_question}} - Vector Store block → query
{{user_question}}, saves Chunk Text into{{context}} - OpenAI / Claude / LLM block → prompt like:
Answer the user's question using only this information:
{{context}}
Question: {{user_question}}
If the information includes a source_url, include it in your answer.Chunks carry their source URL when they came from a crawled website. Telling the AI to "include the source_url if present" gives your bot answers with links to the original page.
In a workflow: the Vector Store block
Found in the workflow editor's Database category. It has two modes:
Search mode (default)
| Setting | What it does |
|---|---|
| Vector Store Name | Pick your vector |
| Query | Search text (variables allowed) |
| Number of Documents | Chunks to return (default 2) |
The result is { statusCode, data: [...chunks] }. Use Save answer with paths like data[0].chunk_text to store results into variables, or pass the whole result to an AI block. Run the single block once to see the Execution Result and pick paths visually.
Upsert mode
Adds one piece of content (with optional JSON metadata) directly into the vector's index.
For building and maintaining a knowledge base, always prefer the Vectors page (add sources + train). Upsert inserts raw entries that bypass the source list — they won't show as sources and won't survive a retrain.
Bring your own Pinecone
If your embeddings already live in Pinecone, skip Indite vectors entirely:
- Workflow: the Pinecone block — search or upsert against your own Pinecone indexes and namespaces, and even create new indexes from inside Indite.
- Bot: the Pinecone Store block in the bot editor's Database section.
Both need a Pinecone API-key credential. The built-in Vector Store needs no credential at all.