Conversations & Citations

Conversations in Opentrace are threaded dialogues within a project where you interact with the AI assistant. Every response is grounded in your documents, with citations linking back to the exact source.

Conversations

Each project can have multiple conversations, allowing you to explore different topics without losing context:

• Create — start a new conversation from the Chat tab
• Switch — jump between conversations in the sidebar
• Delete — remove a conversation and all its messages

Within a conversation, the system maintains a rolling chat history. The last N messages (default: 10) are included in the system prompt to give the AI contextual awareness for follow-up questions like “Tell me more about that” or “What was the second point?”

How Responses Are Generated

1. You send a message
2. The input guardrail validates the message for safety
3. The AI agent queries your knowledge base using the configured RAG strategy
4. Relevant document chunks are retrieved with metadata (document name, page number)
5. The LLM generates a response grounded in the retrieved context
6. The response and citations are returned and stored in the database

Citations

Citations are the cornerstone of Opentrace's trustworthiness. Every piece of information in the AI's response can be traced back to a specific document chunk.

Each citation includes:

• Document name — the source file or URL
• Page number — where the chunk was found (for PDFs)
• Chunk content — the actual text passage used
• Relevance score — how closely the chunk matched your query

Citations accumulate in the agent state across tool calls, so even multi-step agent interactions (in the Supervisor agent) properly track their sources.

Streaming

Responses are streamed token-by-token from the backend to the frontend. You see the answer appear in real-time as the LLM generates it, with citations displayed alongside or below the response once complete.

Message Feedback

Each AI response includes a thumbs-up/thumbs-down feedback mechanism. This feedback is stored in the database and can be used to evaluate and improve the system's performance over time.