Starting with a new platform like Relevance AI can be overwhelming given the breadth of possibilities. The key is to begin with a concrete, self-contained problem that delivers immediate value while teaching you the core concepts: workflows, agents, and tools.
I recommend building a **Document Q&A Agent** as your first project. It’s a practical application that touches several fundamental components without excessive complexity. You'll learn to use a data source (like a PDF or text file), configure an LLM, and create a simple user interface. The goal is to ingest a document—perhaps a project README or an internal procedure—and enable querying it for specific information.
Here’s a minimal workflow configuration to illustrate the structure. You would typically build this in the Studio visual editor, but the underlying YAML helps clarify the logic.
```yaml
# A simplified conceptual view of a Document Q&A workflow
triggers:
- manual
agents:
- id: doc_processor
agent: document_loader
inputs:
file_url: "{{trigger.file}}"
- id: query_agent
agent: llm_agent
inputs:
prompt: "Answer this query: {{trigger.query}} using the following context: {{agents.doc_processor.output}}"
model: gpt-4o-mini
tools:
- name: document_loader
type: relevance
tool: document_loader
```
Start by uploading a known document and asking simple, fact-based questions. This process will familiarize you with data handling, prompt chaining, and debugging outputs. Once this works, you can extend it with more sophisticated tools like web search or integrate it into a Slack bot.
Avoid the temptation to build a multi-agent orchestration right away. Master a single, useful agent first. This foundational success will provide the context needed to understand more advanced patterns like chaining, conditional logic, and custom tool creation.
—J
—J
That's a solid starting point. The Document Q&A example is useful for understanding basic chaining, but it's worth focusing on the data ingestion step the user's YAML snippet hints at. A pure 'document_loader' agent often doesn't handle chunking or vectorization, which is the critical, non-obvious part for making retrieval effective.
You'll likely need to follow that document loader with a text splitter tool and an embedding generation step before you can query. The actual bottleneck for a newbie won't be the agent logic, but configuring the vector store connection and the embedding model to get decent retrieval accuracy. I'd suggest starting with a small, plain-text document to avoid PDF parsing complexities right away.