No default 512 chunk size

Document-type-specific sizing

Rationale for chunk size

Source metadata on chunks

Section/heading metadata

Timestamp or indexed_at metadata

Document preprocessing

Deduplication mechanism

Idempotent re-run design

Markdown-aware chunking

Chunk index/position metadata

Hybrid search implemented

RRF or equivalent fusion

Retrieve-more then rerank fewer

Reranking step present

Not cosine-only

Decoupled embedding model

Empty result edge case

Fusion weighting documented

Design covers all four aspects

Deduplication before rerank

Ingestion pipeline diagram

Retrieval pipeline diagram

Vector DB recommendation with comparison

Chunking parameters specified

Chunking rationale provided

Retrieval metrics specified

RAGAS or equivalent framework

Multi-tenant isolation addressed

Embedding versioning plan

Decoupled embedding adapter

Monitoring plan

Hybrid search in retrieval design

Multiple models benchmarked

Code-appropriate model included

Multilingual model considered

Retrieval metric computed

Comparison table or summary

Model selection stated

Alternative model compared

Versioning or migration noted

Does NOT just default without evidence

Domain relevance addressed

RAGAS framework used

context_precision metric included

context_recall metric included

faithfulness metric included

answer_relevancy metric included

context_precision threshold check

context_recall threshold check

Pass/fail verdict present

Retrieval metrics not skipped

Metric explanations in report

Empty search results handled

Whitespace query handled

Malformed chunk handling in ingestion

Partial batch not silently dropped

Empty results edge case test

Malformed document test

Empty/whitespace query test

Function signatures unchanged

Tenant filter preserved

Deduplication logic preserved