Session
100%
The lab

A hands-on lab for building retrieval you can trust.

Upload a small corpus, then see how your content is analyzed and chunked before retrieval and generation. Every change is measurable, comparable, and reversible, so context stays in focus while your team learns the real tradeoffs.

Analyze structure Chunk with intent Retrieve with context Control enterprise cost

Why this matters: understanding how content is analyzed and chunked can reduce token waste, preserve context, and increase answer reliability so users get content they can trust at a cost the enterprise can afford.

See the tradeoffs live.Tune chunk size, overlap, and retrieval depth in one place, with answer evidence you can inspect immediately.
Evaluate before you ship.Run a compact question set and score faithfulness, relevance, and completeness before committing to architecture choices.
Balance value and spend.Use the cost simulator to keep quality high while storage, compute, and token budgets stay enterprise-affordable.
Ingest Step 1

Start with a small, focused corpus.

Upload a few documents, pick your chunking, and build your first index. This is where retrieval quality is decided.

Upload Documents (max 5 · PDF, TXT, EPUB, DOCX, MD)

📂

Drag & drop files here, or tap to browse

PDF · TXT · EPUB · DOCX · MD · JPG · PNG · WEBP · max 5 files

Pipeline preset ?

See what will run Query-time pipeline for the current selection
How they compare Design-level trade-offs — structural, not measured. Run the Strategy sweep on the Report tab for measured numbers on your corpus.
Preset Ingest $ Query $ Latency Recall Provenance Best-fit query

Reading guide. Latency = end-to-end response time per query at gpt-4o-mini, k=5. Recall = fraction of the truly-relevant chunks that land in the top-k retrieved set (structural pattern from RAG literature; long-context has no ranking so shows n/a). Provenance = fraction of answer quotes that can be verified as substrings of retrieved chunks — the same metric measured on the Report tab. Ranges are structural estimates; run the Strategy sweep for measured numbers on your corpus.

Cost anchors for your corpus Extrapolated at gpt-4o-mini + text-embedding-3-small. Add files above to size these to your corpus.
Preset Ingest (one-time) Per query Per 1,000 queries Assumes

Extrapolated — not measured. Formula is the sum of published per-token prices × structural token counts per preset. Real numbers depend on your queries and the actual model chosen on the Query tab. Contextual Retrieval is added on top when its toggle is on.

How candidate chunks are selected.
Optional reorder of candidates before generation. Rerank cost is tracked separately.
Whether the generator can re-query on low grounding.
Hard chunk cap for long-context mode. Prevents budget surprises.

Every axis choice is recorded in the run's config hash, so the report can be replayed exactly. Presets are lossy summaries — this panel is the source of truth.

Ingestion Parameters

900
Smaller = tighter evidence; larger = richer context.
140
More overlap preserves continuity across related passages.
Use recursive for stable, interpretable chunk boundaries.
Sentence-first favors precise, detail-heavy passages.
Controls how much specific evidence and nuance the answer includes.
Higher values emphasize explicit source quotes and grounded wording.
Choose the search strategy: Compact = tighter focus, Balanced = good default, Deep/Breadth = wider exploration.
Affects embedding cost & retrieval quality
LLM prepends a one-sentence context to each chunk before embedding. Big quality lift; adds one LLM call per chunk at ingest.

Why this matters

  • Smaller chunks can improve precision and reduce context waste.
  • More overlap helps preserve continuity between related passages.
  • Simple defaults usually teach the concept faster than overly complex settings.

What good looks like

  • Fast to build and easy to explain.
  • Grounded answers that cite the right evidence.
  • Clear tradeoffs between relevance, cost, and latency.

Index Ready

Proceed to Query or Report to explore your documents.

Query Step 2

Ask a question that matters.

Start with one prompt that reflects a real user need. Compare how the answer shifts as you change retrieval settings.

Ask a Question

Mode (pre-built demo only)
5
Use 1–3 for tight precision, 5+ for wider coverage
Quick shortcut for retrieval size: Precision uses fewer chunks; Breadth uses more for wider coverage.
0.4
Lower = steadier and more grounded; higher = more creative and less predictable.
1200
0.5
0 = diverse, 1 = relevance-first
0.3
Minimum relevance score to keep a chunk

A simple rule of thumb

  • Lower k usually improves precision and lowers context cost.
  • Higher k can help when the answer needs more coverage.
  • Lower temperature tends to keep the answer steadier and more grounded.

Retrieving and generating answer…

Answer

Sources

Report Step 3

Turn a few questions into an evaluation.

Check whether your setup grounds answers well, and see what to tune next.

Generate RAG Evaluation Report

Add a handful of example questions, run the evaluation, and review the results in a more compact report format.

What this does
You provide a handful of real questions. The lab runs them through your current retrieval setup and grades each answer for faithfulness, relevance, and completeness.
Grounded answers Retrieval quality Chunking insight
How to use it
Choose a few questions that matter to your users, then compare the results across retrieval settings and see which configuration is most reliable.
Before rollout For tuning For demos
5

Executive Summary

Summary

CTO decision drivers

Per-Question Results

Question Faithfulness Relevance Completeness Prov. Latency Cost

k-Sensitivity Sweep

k Faith. Rel. Comp. Avg Latency Cost Tokens

Recommended Next Steps

What to change first
Simulator Step 4

Estimate the cost of scale.

Project storage, compute, and token spend as your corpus grows and content types diversify.

Cost, Compute & Storage Simulator

Adjust the sliders to see how your assumptions change the live estimate. The recommendation panel updates instantly.

Live projection

Cost breakdown

Projected volumes

Corpus profile

Total documents in the target corpus.
Drives text volume and page-level visual parsing.
Text density of a typical page.
Target tokens per text chunk.
Overlap inflates chunk count and embed cost.
Select the data shape that best matches your corpus.

Multimodal content mix

Share of pages that contain each annex type. Higher visual/chart/table/data density shifts the recommended chunking strategy toward layout- and element-aware parsing.

Photos, diagrams, scanned figures → vision captioning.
Plots requiring data-series extraction.
Structured tables → row/column-aware chunks.
CSV/JSON/numeric appendices → record-level chunks.
Model used to caption/extract visual elements.

Models & operations

Drives one-time index embedding cost.
Generation model for live queries.
Context passages per query → prompt tokens.
Avg completion length per answer.
Expected monthly query volume.
Full re-embeds per month (corpus churn).

Recommended chunking strategy

Chunking Step 5

See how chunk structure changes retrieval.

Compare a few chunking strategies on the same passage and inspect how each one reshapes the retrievable surface.

Chunking lab

Use this as a quick mental model for how chunking influences precision, context, and cost.

Drop one file here, or click to upload PDF, TXT, EPUB, DOCX, MD
1 file
No document loaded yet.
Corpus and modality mix
The table below estimates index cost and vector efficiency from this document. Select a row to inspect its actual preview; use the report for measured retrieval quality.
Visual Media Explore

Model visual tokenization for GIS and geodata workloads.

Estimate cost, scale, context relevance, and integrity for visual chunking pipelines inspired by leading GIS and multimodal retrieval practices.

Visual media workload profile

Inspired by ArcGIS multiscale tiling, STAC/COG geodata packaging, H3/S2 spatial indexing, and modern VLM caption plus OCR tokenization flows.

Inputs

What to account for when chunking visual information

Cost, scale, and value signals

Measured backend telemetry comparison

Runs a quick backend report and refreshes measured deltas.

Strategy scorecard

Model provider setup

All four providers are OpenAI-API-compatible. Add whichever keys you want to use to .env, then restart.