Hallucination reduction

Current Workflow

• The user enters a 3GPP RAN/RRC question.

• The primary BM25 retriever searches the local TS 38.331 PDF.

• The system retrieves the top chunks using top_k = 8.

• The system selects up to 18 evidence sentences from the retrieved chunks.

• The generator produces an answer using only the selected evidence.

• The NLI judge checks whether the answer is supported by the evidence.

• If the answer is supported, the system returns the answer.

• If the answer is weak or unsupported, the system runs adaptive retrieval over a fixed list of selected 5G/RRC sources.

• The generator tries again using adaptive evidence.

• The NLI judge checks the adaptive answer.

• If the adaptive answer is still unsupported, the system returns:

• Not enough information.

NLI Judge

• The updated system uses:

• cross-encoder/nli-deberta-v3-base

• The judge compares the generated answer with the selected evidence. It looks at entailment, contradiction, and support.

• The judge can return verdicts such as:

  • supported

  • partially supported

  • unsupported

• This makes the system more evidence-aware because the final answer is not accepted only because retrieval found some related text. The generated answer must also be checked against the evidence.