AI Reasoning
Queria offers full transparency over the AI reasoning process. The reasoning panel shows you exactly how the system analyzes your question, which strategies it adopts and why it selects certain sources.
The reasoning panel
The reasoning panel is an expandable area in the chat interface that appears during the processing of complex queries. It shows in real time every phase of the AI's decision process.
When it appears
The panel activates automatically for questions that require:
- Search across multiple documents or sources
- Comparative analysis between different information
- Multi-step reasoning with logical dependencies
- Integration of certified external sources
For simple questions with a direct answer, the AI replies quickly without showing the full panel.
Reasoning phases
1. Question analysis
The system interprets your question, identifies intent, key entities and complexity level. This phase determines the subsequent retrieval strategy.
2. Planning
The Planner decides the optimal strategy: how many documents to search, which relevance threshold to apply, whether to involve external sources and how to structure the search.
3. Query decomposition
For complex questions, the AI breaks down the request into more specific sub-questions, each solvable independently or sequentially.
4. Document search
The system queries simultaneously company documents, the Knowledge Base and certified external sources, collecting all relevant results.
5. Relevance evaluation
Every result is scored for semantic relevance against the original question. Results below threshold are discarded.
6. Re-ranking
Results are reordered by a specialized model that directly compares each document with the question, improving the final ranking precision.
7. Synthesis
The Writer composes the final answer integrating information from selected sources, adding traceable citations and ensuring consistency and completeness.
Thinking mode
During processing, the AI can show its internal reasoning (chain-of-thought). This is visible in the panel as italicized text and represents the AI's "thinking" as it processes the answer.
Thinking mode is particularly informative to:
- Understand why the AI chose certain sources over others
- Verify that the logical reasoning is correct
- Identify any ambiguities in the question
How to read the panel
- Expandable panel: click to open or close the reasoning detail
- Numbered phases: every step is shown in sequence with a short description
- Execution times: next to each phase is the elapsed time
- Confidence indicators: the system shows how confident it is in the generated answer
- Status icons: completion, in progress or error for each phase
Adaptive complexity
The Cog-RAG architecture automatically adapts the depth of analysis:
| Complexity | Behavior | Documents analyzed |
|---|---|---|
| Simple | Direct answer, few documents | Up to 10 |
| Moderate | In-depth search with re-ranking | Up to 30 |
| Complex | Decomposition and multi-source search | Up to 50 |
| Aggregative | Extended analysis on broad corpus | Up to 200 |
Decomposition strategies
When the question is decomposed, the system adopts one of these strategies:
Sequential
Sub-questions depend on each other. The answer to the first guides the second.
Example: "Who are the heirs in the will and what share goes to each?"
Parallel
Sub-questions are independent and executed simultaneously.
Example: "Compare the contractual conditions of supplier A and supplier B"
Hierarchical
It starts from a general search to then deepen on specific aspects.
Example: "What is the food safety regulation for organic products?"
Comparative
Structured search for direct comparison between two or more elements.
Example: "What are the differences between decree 81/2008 and the new regulation?"
Why transparency matters
- Trust: seeing the reasoning increases trust in the answer
- Verification: you can check that the AI looked in the right places
- Learning: you understand how to phrase future questions better
- Debugging: if the answer is unsatisfactory, the panel shows you where to intervene
TIP
If the reasoning panel shows the AI didn't find relevant documents, try rephrasing the question with more specific terms or activate filters to direct the search.
Queria v3.5.0 -- Cog-RAG Architecture