ORBiS TREE · ENGLISH DOCUMENT
Products →ChatGPT
A conversational AI product from OpenAI that provides natural-language interaction with AI models.
한국어
English
This English page is a curated translation layer linked to the Korean source article. Community changes are currently made on the Korean source, where the full revision history and anonymous edit trail are preserved.
Overview
ChatGPT is a conversational AI service from OpenAI. It was publicly introduced on November 30, 2022. The launch account described follow-up questions, corrections and challenging incorrect premises, while also identifying plausible but incorrect outputs as a limitation.[1]
The model and the service
GPT names a model family; ChatGPT is the user-facing service that connects models with interfaces and available tools. A model's ability does not establish which files or systems a particular account can access. Models, features and limits can change, so record the date, account and environment when describing a capability.
An example of a reviewable request
This is an ORBiS Tree hypothetical example. Instead of asking only for a customer notice, provide confirmed facts about a delayed shipment, specify three paragraphs, prohibit invented dates and ask for the customer's next action. Review the draft for product names, quantities, promises and unsupported benefits. Refine the request with specific corrections. Clear instructions help define the review, but do not guarantee factual accuracy.
Four parts of a request
- Purpose: state whether the task is summarizing, classifying, comparing or drafting.
- Evidence: supply originals and dates; request an explicit unknown when information is absent.
- Output: specify the intended reader and inspectable format.
- Acceptance: check figures, names, sources and prohibited assumptions.
Conversation and external evidence
A long conversation is not a guarantee of complete or current recall. Restate the active goal, confirmed decisions and latest materials when needed. Retrieving relevant parts of larger collections connects to RAG.[3] Important or current claims should be checked against original evidence.
Work and deliverables
ChatGPT Work supports tasks with concrete outputs and multiple steps.[4] Inspect an actual report or spreadsheet for units, source rows, missing periods and computed totals. Keep unsupported causal explanations as hypotheses. Model selection and service access are separate conditions.[5] AI Memory explains stored-state scope, and AI Agent Oversight explains execution and review boundaries.
From supplied material to answers and actions
A document-based question can be divided into interpreting the request, selecting material, generating a grounded answer and reviewing any action. A retrieval failure and a generation failure need different remedies. RAG connects retrieved material to generation. A conversational instruction does not itself retrain model weights; see training and generation.
The Work introduction ↗ is a starting point for specifying goals and deliverables in multistep tasks.[4] When a task opens sources or creates files, inspect actual outputs and their connection to the inputs.
Worked example: a missing return-policy condition
This is a fictional policy, not a store's rules. Document A says ordinary returns must be requested within 14 days of delivery. Document B says items damaged in shipping have a separate verification process. A question about a damaged item received 20 days earlier cannot be answered by applying the ordinary deadline alone or automatically approving an exception. Retrieve the relevant conditions and actual exception procedure.
Record the documents, applicable conditions and remaining unknowns. If the procedure is absent, a grounded answer can state that a separate damage process exists but approval cannot be determined from the supplied sources. Copying the number 14 correctly and applying the rule correctly are separate evaluation items. Distinguish unsupported approval from missing evidence.
Check context, memory and source freshness separately
Check whether the needed policy reaches the next request and whether it has been replaced. Conversation context, persistent information and retrieved originals have different update and deletion mechanisms. Record policy versions, effective dates and product scope with source and transformation history.
If an answer changes, compare inputs, retrieval results, source versions and request conditions before attributing it to the model. Consult the current model guide ↗ for availability in the relevant environment.[5] Switching model names alone does not repair missing context or an incorrect source.
Define completion concretely
An announcement draft may be complete when its conditions are accurate and a readable deliverable exists. Sending it to a customer also requires the recipient, final content and execution result. Permission to send and confirmed delivery are separate facts. Continue to action oversight for checkpoints and duplicate prevention, and the work environment for file and analysis completion criteria.
SOURCES
Sources
-
Introducing ChatGPTOpenAIOpen source ↗
-
Prompt engineering best practices for ChatGPTOpenAI Help CenterOpen source ↗
-
Retrieval-Augmented Generation for Knowledge-Intensive NLP TasksLewis et al. / NeurIPSOpen source ↗
-
Get started with WorkOpenAIOpen source ↗
-
Models in ChatGPTOpenAIOpen source ↗
KNOWLEDGE LINKS
Continue from here
BACKLINKS