Agentic Workflows: composable RAG pipelines
vai v1.25.0 introduces agentic workflows: multi-step RAG pipelines defined as portable JSON files. Search multiple collections, merge results, filter by relevance, and summarize with LLMs, all from a single command.
The problem
Individual vai commands handle single operations well. But real RAG pipelines often need multiple steps chained together:
- Search two collections, merge results, then rerank
- Check if a document already exists before ingesting
- Query a knowledge base, then summarize results with an LLM
Previously, you would write a bash script or Node.js code to chain these operations. Workflows let you define the pipeline declaratively.
How it works
A workflow is a JSON file that defines a directed acyclic graph of steps. Each step maps to a vai operation and can reference outputs from previous steps using template expressions:
{
"name": "Multi-Collection Search",
"steps": [
{
"id": "search_api",
"tool": "query",
"inputs": {
"query": "{{ inputs.query }}",
"collection": "api_docs"
}
},
{
"id": "search_arch",
"tool": "query",
"inputs": {
"query": "{{ inputs.query }}",
"collection": "architecture"
}
},
{
"id": "merge",
"tool": "merge",
"inputs": {
"arrays": [
"{{ search_api.output.results }}",
"{{ search_arch.output.results }}"
]
}
}
]
}
vai analyzes step dependencies and runs independent steps in parallel. In the example above, the two searches run simultaneously, then merge waits for both to complete.
Built-in templates
vai ships with 5 templates covering common patterns:
| Template | Description |
|---|---|
multi-collection-search | Search multiple collections, merge, rerank |
smart-ingest | Deduplicate before ingesting |
research-and-summarize | Search then summarize with an LLM |
consistency-check | Compare content across collections |
cost-analysis | Compare embedding costs across models |
Try one:
vai workflow list
vai workflow run multi-collection-search \
--input query="how do embeddings work" \
--input collection1=api_docs \
--input collection2=tutorials
No auth, no accounts, no hosted services
Workflows are JSON files that live in your project. No accounts, no cloud services, no lock-in. Commit them to git, share with your team, run them anywhere vai is installed.