Guide
Understanding Responses
Learn how AI processes your questions and provides answers with thinking stages and citations.
When you chat with AI in Onsomble, you see more than just the final answer. You see the AI’s reasoning process, source citations, and metadata about the response.
The Three Lanes
Every AI response has three parts:
| Lane | What It Shows |
|---|---|
| Thinking | Processing stages — what the AI is doing |
| Content | The actual response text |
| Artifacts | Citations, model info, and actions |
During Streaming
While the AI is generating a response:
- Thinking lane shows live progress through stages
- Content lane streams text as it’s generated
- Artifacts lane accumulates citations as they’re found
After Completion
Once the response is complete:
- Thinking collapses to “Thought for X seconds” (expandable)
- Content shows the full response with citation links
- Artifacts shows the model used, source count, and action buttons
Thinking Stages
Thinking stages show what the AI does to answer your question.
Common Stages
| Stage | What’s Happening | Notebook | General |
|---|---|---|---|
| Classification | Analyzing your question type | Yes | Yes |
| Vector Search | Finding relevant passages | Yes | No |
| Graph Search | Traversing knowledge graph | Yes | No |
| Reranking | Sorting results by relevance | Yes | No |
| Web Search | Searching the internet | Optional | Optional |
| Generation | Creating the response | Yes | Yes |
Reading Stage Details
Click on any stage to see details:
Classification details:
- Query type (factual, analytical, comparison, etc.)
- Confidence score
- Whether retrieval is needed
Vector Search details:
- Number of results found
- Similarity scores
- Which sources were searched
Rerank details:
- How many results were considered
- Score improvements
- Cache hit (faster if recently searched)
Why Stages Matter
Thinking stages help you:
- Understand the process — See how the AI arrived at its answer
- Debug issues — If retrieval found nothing, you know to add better sources
- Build trust — Transparent reasoning shows the AI isn’t making things up
If you see “0 results” in vector search, your sources might not contain relevant content for that question.
How Citations Work
Citations connect AI responses to their sources.
Citation Format
In the response text, you’ll see numbered markers like [1], [2], [3].
“The study found a 25% improvement in outcomes [1], consistent with earlier findings [2].”
Citation Types
| Type | Source | Icon |
|---|---|---|
| RAG | Your uploaded documents | File icon |
| Web | Internet search results | Globe icon |
Viewing Citations
Quick preview: Hover over a citation number to see a popover with:
- Source title
- Relevant excerpt
- Link to full source
Full details: Click “Sources” in the message footer to open a dialog showing all citations with:
- Complete excerpts
- Source metadata
- Direct links
How RAG Citations Are Generated
- Your question triggers a semantic search
- Relevant passages are retrieved from your sources
- The AI uses these passages to generate an answer
- Each passage becomes a numbered citation
- Citations link back to the original source and location
The RAG Process Explained
RAG (Retrieval Augmented Generation) is how notebook chat grounds answers in your sources.
Query Classification
The AI analyzes your question:
- What type of question is this?
- What kind of answer is expected?
- What retrieval strategy should be used?
Semantic Search
Your question is converted to a vector (a numerical representation of meaning). This vector is compared against all the passages in your sources.
Passages with similar meanings score higher — even if they don’t share exact words.
Knowledge Graph (Optional)
If enabled, the AI also searches your notebook’s knowledge graph for related entities and relationships.
Reranking
The initial results are re-scored using a more sophisticated model. This improves relevance by considering:
- How well the passage answers the question
- The passage’s position and context
- Redundancy with other passages
Context Assembly
The top passages are assembled into a context window for the AI. Each passage includes:
- The text content
- Source information (for citations)
- Relevance score
Generation
The AI reads the context and generates a response. It’s instructed to:
- Only use information from the provided context
- Cite sources for each claim
- Indicate when information isn’t available
Response Metadata
After each response, you can see metadata about how it was generated.
What’s Included
| Field | Description |
|---|---|
| Model | Which AI model generated the response |
| Credits | How many credits were used |
| Duration | How long generation took |
Viewing Metadata
Click the info icon on any AI message to see:
- Token usage (input and output)
- Credit cost breakdown
- Model information
Message Status
Messages go through several states:
| Status | Meaning |
|---|---|
| Streaming | Response is being generated |
| Completed | Response finished successfully |
| Error | Something went wrong |
| Approval Pending | Waiting for your approval (agentic workflows) |
Error Messages
If something goes wrong, you’ll see an error message explaining what happened. Common issues:
- Credit insufficient — You need more credits
- Rate limit — Too many requests, wait and try again
- Timeout — Request took too long, try simplifying
Agentic Workflows
For complex questions, the AI may create a plan and ask for your approval.
How It Works
- You ask a complex question
- AI generates a research plan
- You see the plan with an Approve/Reject choice
- If approved, AI executes the plan
- Results stream back with detailed progress
When This Happens
Agentic workflows trigger for:
- Multi-step research tasks
- Complex comparisons
- Deep analysis requests
You can approve, modify, or reject plans. Rejecting is free — you only use credits when work is performed.