L0S2 · Data & RetrievalS6 · Security & FinOpsVerified: 2026-11-17
Qdrant Dense + SPLADE Hybrid Search & ACL Pre-Filtered HNSW Primitive
Vector retrieval primitive combining dense embeddings and learned sparse lexical vectors (SPLADE/BM25) via Reciprocal Rank Fusion while enforcing tenant ACL pre-filtering directly inside filterable HNSW traversal.
Best For
Enterprise knowledge bases mixing exact identifiers (error codes, SKUs, clause numbers) with semantic paraphrase queries under strict tenant RBAC isolation.
Human-AI Boundary
Document ACL mutations from upstream IAM systems drive deterministic payload updates; the retrieval engine enforces explicit allowlists without guessing permissions.
Anti-Patterns to Avoid
- Applying RBAC tenant/department filtering AFTER top-K HNSW search (post-filtering), which silently returns 0 results when a restrictive tenant owns <1% of matching chunks
- Relying solely on dense embeddings for alphanumeric part numbers, ticker symbols, or regulatory form IDs where exact lexical match is mandatory
Latency/Cost Profile & Eval Metrics
Hybrid dense + sparse + RRF pre-fetch < 35ms P95 on 5M vectors per shard · RAM footprint for HNSW graph + scalar/binary quantization trade-offs
- Recall@20 under Selective ACL Filters (1% selectivity): >= 0.92 with 0 unauthorized tenant chunk leakage (Audit across minority tenants holding <1% of corpus volume)
- Exact-Token MRR@10 (Code/Clause Queries): +15% lift over dense-only retrieval (Separate evaluation slice for error codes, acronyms, and numeric identifiers)
💰 Business ROI Model & Evidence Level
verified_benchmark📊 Benchmark & Alternative Comparison
- Qdrant Dense + SPLADE + Pre-Filtered HNSW: unified single-cluster storage with segment-aware ACL pre-filtering and native binary quantization.
- Elasticsearch/OpenSearch Hybrid: mature BM25 inverted index strength paired with HNSW, heavier JVM cluster operations for high-dimensional quantized dense vectors.
⚡
Test muscle memory on this topic via 3-Layer Flashcards
Active Recall: Mentally formulate Prompt ➔ Math Derivation ➔ Benchmark Implementation in 3 minutes.
Bidirectional Taxonomy Links
[L1] Stage 2 · Incremental CDC Document Sync, Content-Hash Dedup & Tombstone GC Pattern[L1] Stage 2 · Semantic-Layer Grounded Text-to-SQL & Read-Only Sandbox Repair Pattern[L2] L2 Industry Playbook · E-Commerce Multimodal Search, Attribution & Guided Buying AI[L2] L2 Scenario · E-Commerce Multi-Policy Returns & Claims Adjudication Copilot[L3] Capstone: Multi-Tenant RBAC, Incremental CDC & Semantic Cache RAG Hub[HubA] Company Intelligence Hub · Google DeepMind, Gemini & Cloud AI Loop[HubB] Role Lens · Forward Deployed Engineer (Enterprise AI Delivery & Solutions)[L0 Card] Hybrid Search (BM25 + Dense) & RRF[L0 Card] BM25 Sparse Retrieval[L0 Card] HNSW Multi-Layer Skip Graph[L0 Card] Dense Bi-Encoder Retrieval
Internal Tools & Primary Sources
- Qdrant Hybrid Queries & Filterable HNSW Documentation (official_docs, 2026-11-17)
- SPLADE: Sparse Lexical and Expansion Model for First Stage Ranking (paper, 2026-11-17)