feat: Add normalized metadata reranking and query expansion (Fixes #40)

[zohaib-7035]

Summary

This PR re-introduces the Metadata Reranking feature and Query Expansion requested in issue #40, but completely rewrites the mathematical scoring logic to strictly address the maintainer feedback regarding distorted search results and flat variable additions.

Changes

• Log-Normalization: Applies math.log10() to citation counts to safely compress massive outliers (e.g. 10,000 citations).
• Min-Max Scaling: Scores for years and citations are scaled proportionally from 0.0 to 1.0 based on the dataset batches.
• Strict Bounded Multipliers: Instead of doing flat unrestricted addition (e.g. + 10.0), we apply max-capped multipliers: Citations (max 1.15x), Year (max 1.10x), Trusted Source (max 1.05x). This strictly restricts the absolute maximum possible metadata boost to + 30%.
• Semantic score quality can no longer be "drowned out" by metadata. A weak vector match will remain weak, preserving semantic/keyword query relevance heavily.
• Re-added the synonym mapping for Query Expansion (mouse brain, eeg, fmri, etc).
• Included completely new strict Pytest unit tests in test_metadata_rerank.py mathematically proving the boundary multiplier limits are protected safely.

Test Output Proof (Passing Screenshots)

Notice how the multipliers are mathematically bounded and all Pytest boundaries pass cleanly!

============================= test session starts =============================
platform win32 -- Python 3.12.10, pytest-9.0.2, pluggy-1.6.0
collected 3 items

backend/tests/test_metadata_rerank.py::test_rerank_max_bounds PASSED     [ 33%]
backend/tests/test_metadata_rerank.py::test_rerank_log_normalization PASSED [ 66%]
backend/tests/test_metadata_rerank.py::test_rerank_empty_metadata_handling PASSED [100%]

============================== 3 passed in 0.28s ==============================

[QuantumByte-01]

The reranking logic is sound (bounded multipliers, log-normalization) and the test structure is good. Three issues to fix:

1. Factual error in QUERY_SYNONYMS (critical)

"mouse brain": ["Rattus norvegicus", ...]

Rattus norvegicus is rat, not mouse. Mouse is Mus musculus. Expanding a 'mouse brain' query to return rat datasets would give users wrong results. Fix the mapping.

2. expand_query is only wired into smart_knowledge_search
smart_knowledge_search is only called when filters are provided. The primary search paths (general_search, general_search_async) don't use expand_query. Either apply it consistently across all search entry points, or add a comment explaining why it is intentionally selective.

3. Float precision risk in tests

assert ranked[0]["_rerank_multiplier"] == 1.30  # can be 1.3000000000000003
assert ranked[0]["_score"] == 130.0

IEEE 754 float arithmetic (1.0 + 0.10 + 0.15 + 0.05) can produce 1.3000000000000003. Use pytest.approx:

assert ranked[0]["_rerank_multiplier"] == pytest.approx(1.30)
assert ranked[0]["_score"] == pytest.approx(130.0)

[zohaib-7035]

Hi @QuantumByte-01 , I’ve addressed all three requested changes:

• Fixed QUERY_SYNONYMS replaced "Rattus norvegicus" with "Mus musculus" under the "mouse brain" key. Good catch!
• Wired expand_query consistently across general_search(), general_search_async(), and smart_knowledge_search() so query expansion works everywhere.
• Updated float assertions in test_metadata_rerank.py to use pytest.approx() for better handling of precision edge cases.

All tests are now passing. Ready for your re-review-let me know if anything else is needed. Thanks!