Semantic Retrieval
Vector-first retrieval for concept-level matches across long enterprise content.
When to use: Use for knowledge-heavy workloads where meaning is more important than exact keyword match.
Tradeoff: Higher recall, may need reranking to improve precision.
Hybrid Retrieval
Combine semantic retrieval with lexical search for stronger recall and precision.
When to use: Use for legal, compliance, and reporting flows that require both intent and exact term coverage.
Tradeoff: More configuration and runtime complexity, better groundedness outcomes.
Reranking and Cross-Encoder Stage
Apply a second-stage relevance ranker to tighten top-k evidence quality.
When to use: Use for customer-facing and compliance-critical outputs where precision matters most.
Tradeoff: Adds latency/cost, significantly improves relevance quality.
Query Understanding
Rewrite, decompose, and classify queries before retrieval execution.
When to use: Use for ambiguous, multi-part, or domain-specific user questions.
Tradeoff: Extra orchestration step, improves retrieval targeting.
Groundedness Controls
Attach citation and grounding policies to generation and approval flows.
When to use: Use when outputs are compliance-sensitive, auditable, or externally published.
Tradeoff: Can reduce generation flexibility, improves trust and auditability.
Reciprocal Rank Fusion (RRF)
Merge ranked lists from multiple retrieval sources using reciprocal rank scoring.
When to use: Use when combining dense and sparse retrieval signals for higher recall.
Tradeoff: Simple fusion, no trainable parameters — may not capture nuanced relevance.
Adaptive Strategy Selection
Automatically select the best retrieval strategy per query based on performance history.
When to use: Use when query types vary widely and no single strategy dominates.
Tradeoff: Requires performance tracking data; cold-start defaults to hybrid.
Hypothetical Document Embeddings (HyDE)
Generate a hypothetical answer, embed it, and use the embedding to retrieve real documents.
When to use: Use for abstract or exploratory queries where direct embedding yields poor recall.
Tradeoff: Extra LLM call per query increases latency and cost.