Grounded AI: making an LLM draft a project charter from real data — without hallucinating

Описание и применение

Автоматизированный рецепт генерации проектных чартеров с использованием LLM, основанный на реальных данных, без галлюцинаций.

Задача

Project charters often contain fabricated information when generated by LLMs, leading to unreliable documents in governed PM tools.

Результат

The charter generator now produces accurate, non-hallucinated prose sections, respects existing data, and outputs deterministic JSON for easy integration.

Рабочий процесс

• 1. Collect required input documents (business case, benefits plan, agreements) and cap each to 8,000 chars, total 40,000 chars.
• 2. Build grounding context with deterministic facts and extracted prose excerpts.
• 3. Define a system prompt that (a) grounds all sections on provided data, (b) refuses to fabricate missing data, and (c) returns JSON with only specified keys.
• 4. Create a user prompt that lists target section keys and injects the grounding context.
• 5. Send prompts to the LLM and parse the JSON response.
• 6. Apply the generated prose only to empty sections, skipping live-data tables and preserving existing content.
• 7. Store the result in a draft state for human review and approval.
• 8. Ensure the AI path is optional (BYOK) and fails safely if not configured.

Доказательства

The model will fabricate unless you make refusing the easier path.

Источник