The position-weighting heuristic in sdk/nexent/core/nlp/tokenizer.py:52 compares a token index (idx) against one-third of the input character length:

position_factor = 1.2 if idx < 3 or idx > len(text) / 3 else 1.0

The intent — based on the comment one line above — is "words at the beginning and end of the sentence are more important". But len(text) / 3 is computed against the raw character string, not the number of tokens. For any reasonable Chinese query of more than ~10 characters, len(text)/3 is already larger than the actual number of tokens, so the idx > len(text)/3 branch never fires and only the first three tokens get the boost. For short queries the opposite happens: nearly every token gets boosted.

Repro

from nexent.core.nlp.tokenizer import calculate_term_weights
# 30-char sentence with ~10 tokens; len(text)/3 == 10 �� "end-of-sentence" boost never applies
print(calculate_term_weights("人工智能在医疗影像识别中的应用与前景分析"))

Expected

position_factor should be a function of position within the token stream, e.g.:

n_tokens_estimate = max(1, len(text) // 2)  # or pre-tokenize first
position_factor = 1.2 if idx < 3 or idx >= n_tokens_estimate - 3 else 1.0

A cleaner fix is to materialize pseg.cut(text) into a list first and use len(words) as the denominator.

Severity

Medium — search ranking quality is degraded but the function still returns something usable.