Adopt a retrieval-first workflow: retrieve a bounded set of source passages, preserve mechanics and numbers, use low temperature and a token limit, and reject answers that cannot cite source chunks.

Operate on a verified, OpenAI-compatible surface with environment-based API keys, explicit error handling, and retry/backoff strategies (including 429 handling) to ensure resilience.

Log detailed request and processing metadata (request ID, input size, chunk IDs, model status, output length) to enable explainable rollbacks without exposing private data.

Introduce a Node.js integration approach for OpenAI GPT API text summarization with a strict quality gate ensuring grounded, source-cited outputs for player-facing answers.

Measure SLOs via p95 latency across retrieval, model computation, and application layers; monitor omission and grounding to avoid missing critical game mechanics or values.

Provide a safety-focused integration pattern: prompt should declare operation, audience, and output contract; instruct the model to admit when the source cannot answer.

Reference official API guidance and related vector/search tooling to support a private knowledge base setup.

Choose a backend boundary based on needs: direct OpenAI for unified retrieval with control, or a unified REST boundary to support multiple runtimes and languages, depending on organizational constraints.

For small corpora, direct summarization may suffice; for larger corpora, implement chunking and retrieval to ensure correctness and traceability.

Maintain a robust verification and rollback process: keep an evaluation set of real questions, assess groundedness and omission separately from latency metrics, and implement feature flags to rollback if groundedness declines or budgets are breached.

Provide specific implementation pointers: how to structure prompts, the role of temperature and max_tokens, and Go and Node.js integration patterns, including a Go sample for safe, restricted chat-completions.

Treat truncation as a quality boundary; prefer splitting sources into chunks and summarizing each chunk, then aggregating, rather than truncating relevant passages.