Improving ChatGPT Real‑time Accuracy with Document Retrieval: A Practical Approach

1. ChatGPT Usage Issues

Real‑time Latency

ChatGPT is trained on data up to February 2021, so its answers can be outdated; when used as a search engine it cannot provide current information.

Example screenshots show a 2021‑02 answer for global population ranking, illustrating the time limitation.

Questionable Accuracy

An example from MySQL InnoDB vulnerability demonstrates ChatGPT giving a plausible but incorrect answer, which could mislead users who trust the response.

2. Proposed Solution

Method Overview

Use document search and machine reading comprehension to retrieve up‑to‑date references, then feed the combined prompt to a large language model (e.g., GPT‑4) to generate accurate answers.

This shifts the burden from the model’s internal knowledge to an external retrieval component, improving both timeliness and correctness.

Example Workflow

Step 1: Query the user’s question (e.g., “CVE‑2021‑2429 vulnerability fix”) and retrieve relevant documents.

Step 2: Combine the retrieved text with the original question into a prompt.

Step 3: The LLM answers based on the provided material; multi‑turn dialogue remains intact.

Images illustrate the flowchart and each step:

3. Conclusion

ChatGPT’s errors stem from static training data; by augmenting it with real‑time document retrieval, its responses become more reliable, similar to how humans consult references to reduce mistakes.