RAG Performance Dashboard

real embeddings This is the anatomy of a retrieval-augmented answer. Pick a question, then move the levers on the left — the same ones tuned in production. Document scores and the retrieval-quality metrics are measured: the corpus is genuinely embedded (three models, three enrichment levels, computed offline) and the cosine math runs in your browser. Latency, tokens, and cost are modeled from published pricing. The LLM answer is pre-authored per quality tier — no API calls leave this page.

Chunking

Boundary Strategy

Chunk Size (tokens)

Chunk Overlap

Enrichment Level (what gets embedded)

Embedding Model (all computed for real, offline)

HyPE (index-time question gen) (no query cost)

Index Quantization

Chunk Preview

Index Options

Contextual Retrieval (LLM context at index time)

Query Enhancement

HyDE (+300ms, extra LLM call)

Query Expansion Variants

Intent Classification (type routing)

Query Decomposition (compound queries)

Step-Back Prompting

Retrieval

Retrieval Mode

Initial Top-K

Score Threshold 0.00

MMR Diversity

Reranking

Reranking Model

Section Context Prepend

Lost-in-Middle Reorder (free win)

LLM + Context

Provider

Model

Temperature 0.20

Final Top-K to LLM

Adaptive Token Budget (−28% tokens)

Context Compression (−45% tokens, −7% quality)

Caching

Exact Embedding Cache

Semantic Cache

Prompt Caching (provider feature)

Presets

Retrieved Documents

Pipeline Trace

Answer

Raw LLM Interaction

System Prompt

Retrieved Context

User Query

LLM Response

LLM Provider

Retrieval Quality · measured

vs ground-truth relevance labels, this query, current pipeline

Embedding Lab · measured

Similarity of all 27 chunks to this query — relevant chunks in sage, noise in slate. Separation is what retrieval quality looks like.

cosine similarity

Enrichment lift — corpus-average MRR by what gets embedded:

Model comparison — corpus-average, current enrichment:

Quality (RAGAS-style) · modeled

Latency · modeled

0

Embed

Retrieve

Rerank

LLM

Other

Token Consumption · modeled

System prompt
Retrieved context
Query + enhancements
Output
Cache discount
Total

Cost · modeled

Per query

Per 1,000 queries