fix(retriever): avoid re-normalizing full in-memory matrix

[UdayKumar9381]

Description

Closes #27.

This PR optimizes InMemoryRetriever.add() so that incremental ingestion does not rebuild and re-normalize the entire stored vector matrix on every call.

Changes

• Normalize only newly added vectors.
• Store newly normalized vectors as matrix blocks.
• Concatenate matrix blocks lazily on the first retrieval after new data is added.
• Preserve existing retrieval behavior and result ordering.
• Add a regression test comparing single-batch and multi-batch ingestion.
• Update CHANGELOG.md under [Unreleased] -> ### Fixed.

Validation

• pytest tests/test_retriever/test_in_memory.py -q — 33 passed
• pytest -q — 266 passed
• ruff check ragframework/ tests/ — passed
• pre-commit run --files CHANGELOG.md ragframework/retriever/in_memory.py tests/test_retriever/test_in_memory.py — passed
• git diff --check — passed

Type of change

• Performance improvement
• Bug fix
• Tests added/updated
• Documentation/changelog updated

[adaumsilva]

@UdayKumar9381 Thanks for the optimization! Tests and CI pass, and retrieval ordering remains correct.

One issue needs fixing before merging: after concatenation in retrieve(), _matrix_blocks still retains every original array alongside the new full matrix. This doubles persistent NumPy vector storage.

Please consolidate the blocks so they reference the cached matrix after concatenation, and reuse the existing array when there is only one block. Add regression coverage confirming that consolidation avoids duplicate storage and that subsequent add/retrieve cycles preserve all results.

Please request another review once these changes are pushed and CI passes.

[UdayKumar9381]

Thanks for the review.

I’ve addressed the requested changes:

Consolidated _matrix_blocks after retrieval to avoid duplicate NumPy storage.
Reused the existing array when only one matrix block is present.
Added regression coverage for consolidation and subsequent add/retrieve cycles.
Verified the full test suite and Ruff checks pass.

The changes have been pushed to the PR. Please take another look when convenient.