fix: prioritize exact numeric knowledge retrieval

[EDDIWARD]

Summary

Prioritize exact numeric identifiers in knowledge-base retrieval. Numeric queries now include an exact phrase score and use a higher accurate-search weight only when the caller has not set one.

Root cause

The semantic score can make nearby numeric identifiers look similar. In the original implementation, a full-width identifier such as １０．１２８．０．４２ was not kept as one numeric token; the record with １０．１２８．０．５２ could therefore rank first.

Before

Querying １０．１２８．０．４２ ranked the near-but-wrong record containing １０．１２８．０．５２ first.

After

The same query ranks the exact １０．１２８．０．４２ at 100.0% relevance.
wrong record containing １０．１２８．０．５２ at 0.9% relevance

Validation

• Focused backend and SDK tests: 378 passed.
• Manual local reproduction with the same full-width numeric query: the exact record is ranked first.
• No migrations, configuration changes, or deployment files are included.

Screenshots

The first image is the reproduction before the fix; the second image is the fixed result.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!