FDE Foundations · Module 7: RAG & Retrieval
Hybrid Search That Works
Keyword search finds exact terms and codes; vector search finds paraphrases. Hybrid with a fusion step beats either alone for most enterprise corpora.
10 min reading
Objectives
- Combine keyword and vector retrieval with a fusion step
- Tune the top-k and filtering strategy
- Explain recall vs precision in retrieval terms
Why hybrid
Enterprise queries contain exact tokens: invoice numbers, product codes, error messages. Pure vector search blurs those; pure keyword search misses phrasing. Run both, then fuse result lists (reciprocal rank fusion is a standard, simple choice) and re-rank the merged list if the budget allows.
Top-k and filters
Retrieve enough candidates to survive fusion: top 20 to 50 from each system, filter by metadata before fusion (access control, document type, date), then keep the best few for the answer step. Filtering before ranking is also your permission enforcement point; do it in the retrieval query, never after, or revoked documents leak into answers.
Recall vs precision, in the field
Recall: did the right document enter the candidate set at all? Precision: is the top of the list the right one? Triage failures accordingly: the answer missed the fact entirely points at recall (chunking, index coverage); the answer used the wrong document points at ranking (fusion weights, re-ranker).
Measure before tuning
Build a small evaluation set of real questions with known target documents. Every tuning change runs against it. Tuning without that set is guessing with extra steps.
Quick check
An optional 2-3 question self-check. Answers never leave your device, are not stored, and never count toward any assessment.
Exercise
Design retrieval for a fictional support knowledge base: the two retrieval methods, fusion approach, top-k numbers, the filter applied before ranking, and two example queries showing why hybrid wins.
Pass criteria
Both methods named, fusion approach stated, numbers concrete, the pre-ranking filter includes access control, and the two examples show a keyword win and a semantic win.