Assistente do SEI IA e Engenharia de Prompts

The video titled "Assistente do SEI IA e Engenharia de Prompts" is a comprehensive workshop focused on the use, development, and best practices of an AI assistant integrated with the SEI (Electronic Information System) used at Anatel, a Brazilian regulatory agency. The main technological concepts, product features, and Prompt Engineering techniques covered include:

Key Technological Concepts and Product Features

• CIA (Assistente do SEI IA):
  • An AI assistant integrated within SEI to facilitate interaction with documents and processes, respecting business rules and access controls.
  • Uses GPT-4.1 and a "reflection" model (O4 mini) for more precise, context-aware responses.
  • Does not search the web online yet; a web search feature and personalized GPT agents are planned for future releases.
  • Designed for corporate and secure use, avoiding internet-based AI tools for sensitive or restricted content to prevent data leaks and unauthorized use of proprietary data.
• User Interface and Features:
  • Assistant located at the bottom right corner of the SEI editor.
  • Includes chat interface, icebreakers (predefined simple prompts), favorites, prompt galleries (shared and personal), and topic management (archiving/unarchiving).
  • Supports citation of documents and processes with context limits (up to 1 million tokens).
  • Incorporates vector search technology (HAG - coarse search sieve) to locate relevant document excerpts before AI processing.
• Security and Compliance:
  • Emphasizes strict data privacy and contractual protections.
  • Warns against using free internet AI tools for corporate content due to data usage policies and security risks.

Prompt Engineering and Usage Guidelines

• Prompt Definition and Importance:
  • A prompt is any input message to the AI, from simple greetings to complex instructions.
  • Effective Prompt Engineering is crucial to get accurate, relevant, and safe AI outputs.
• Prompt Types and Techniques:
  • Zero-shot: Asking questions without examples.
  • One-shot: Providing one example to guide the AI.
  • Few-shot: Providing multiple examples.
  • Chain of Thought: Asking AI to think step-by-step or reason through a problem explicitly.
  • Persona-based prompts: Assigning a role or expertise to the AI to tailor responses (e.g., productivity coach, tax lawyer).
  • Task decomposition: Breaking complex tasks into smaller subtasks to improve accuracy.
  • Self-reflection/self-criticism: Asking AI to evaluate its own responses for quality and correctness.
  • Hypothetical search: Creating fictitious references or theses to guide research in official databases.
• Prompt Writing Best Practices:
  • Start simple and evolve prompts gradually based on feedback.
  • Assign a clear persona with role, expertise, and context.
  • Provide detailed context, target audience, and clear objectives.
  • Specify restrictions explicitly (e.g., "do not invent," "base exclusively on referenced documents").
  • Use imperative or infinitive verbs for clarity and directness.
  • Include positive and negative examples to optimize AI responses.
  • Specify response format, length, tone, and style (Markdown formatting recommended).
  • Avoid ambiguity by using synonyms and clear terms (e.g., distinguish between listing and transcribing).
  • Use Markdown and XML-like delimiters to structure outputs clearly, especially for complex prompts.
  • Regularly test and refine prompts based on results.
• Advanced Prompt Templates:
  • Provided in the manual and ready-made gallery, including complex legal and administrative document analysis.
  • Examples include detailed administrative appeal analysis with explicit instructions on format, content, and style.
  • Use of tables, bold/italic formatting, and structured sections to improve readability and precision.

Practical Applications and Examples

• Summarizing texts with topic limits and emojis for social media sharing.
• Translating texts into multiple languages (French, English, Spanish).
• Programming assistance (Python, SQL, Excel formulas).
• Legal document analysis, including citation and cross-referencing of reports and decisions.
• Risk analysis from news clippings, associating news items with institutional risks, generating tables with justifications.
• Persona-driven debates simulating opposing roles (e.g., lawyer vs. prosecutor) to refine arguments and responses.

Challenges and Limitations

• AI is literal and probabilistic, not truly understanding content; it can hallucinate or produce nonsense if prompts are unclear or incomplete.
• The AI assistant cannot telepathically guess user intent; specificity is required.
• Document size and token limits require breaking large documents into smaller parts.
• Managing access rights and ensuring the AI only processes information accessible to the user’s unit.
• Avoiding data leaks and ensuring compliance.