01
Point at your data
Start with local text, documents or a folder. No prepared workspace required.
Files and schemasv0.2.1New HTML reports. Same inspectable evidence.
Local files. Controlled experiments. Original evidence.
See whether your RAG change helped.
Compare retrieval techniques on your own data. Inspect what changed before bringing it into your RAG project.
Install CLI · Python 3.12
uv tool install --python 3.12 "ragvidence==0.2.1"Runs locally · BM25 needs no model or API key
No account
No telemetry
No hosted database
Your model, or none
01 / THE WORKFLOWFROM COMMAND TO EVIDENCE
Work on your own data
Start with a question.
Follow the evidence.
Compare alternatives locally. Inspect original sources, returned context and resource trade-offs.
Bring your own dataSYNTHETIC PREVIEW
Four reviewed dev questions. Three retrieval techniques. All outcomes remain visible.
From command to evidence
Synthetic own-data previewSTEP 01 / 04
One command. Your own machine.
Terminal
uv tool install --python 3.12 "ragvidence==0.2.1"Install the CLI in an isolated tool environment.
BM25 and the synthetic example need no account or API key.
Step 1 of 4: Install. One command. Your own machine.
02 / THE EXPERIMENTONE FROZEN INPUT. USEFUL ALTERNATIVES.
Inspect first. Review when you need scores.
From your files to a retrieval decision.
Inspect your sources, compare retrieval, then review the evidence.
02
Compare alternatives
Change retrieval, context or ranking. Keep data and questions frozen for a useful comparison.
Techniques and requirements03
Inspect the evidence
Read returned context and resource use. Reviewed references reveal gains and regressions.
Read your first reportTerminal
ragvi compare-data ./docs --questions ./questions.txt --techniques bm25,parent-childQuestions alone remain unscored. Add reviewed evidence when you need quality metrics.
03 / THE WALKTHROUGHREAL COMMANDS. INSPECTABLE RESULTS.
Your questions. Your evidence.
Follow an experiment all the way through.
Compare BM25, parent-child retrieval and sentence extraction on a fictional device manual, then inspect returned evidence in the actual report.
Silent workflow: ask a question about local data, compare techniques, export HTML, JSON and Markdown, then inspect original evidence. Click or press Space or Enter to pause or play. Right-click for browser video options.
34-second workflow · Synthetic manual · 4 reviewed dev questions · k=1 · CLI 0.2.1.
Ask a question, compare techniques and export HTML, JSON or Markdown. Execution waits shortened; results are specific to this mock dataset. Click or press Space to pause/play. Right-click for browser video options.
04 / THE EVIDENCESYNTHETIC EXAMPLE · DEV ONLY
A trade-off you can inspect
The evidence improved. The context grew.
A fictional device manual exposes a missing power condition. Returning the full parent recovers it, while returning more text.
All runs use four reviewed dev questions, k=1, 100-character chunks and 10-character overlap. Sentence extraction leaves the same question insufficient.
Open the full reportPyPI 0.2.1 · reviewed original spans · no model
| Technique | Sufficient | Mean context¹ |
|---|---|---|
| bm25 | 3 / 4 | 161 chars |
| parent-child | 4 / 4 | 741 chars |
| compress | 3 / 4 | 200 chars |
¹ Complete rendered context, including IDs and headings. All results are shown. This is teaching data, not a benchmark or heldout evaluation.
Q1 / RESET + POWER
The short BM25 hit contains the reset duration. The returned parent also contains the required USB-C power condition.
Four questions cannot establish a general best technique. The CLI does not judge generated answers, citations or abstention.
05 / OPEN SOURCEALL SHIPPED FEATURES INCLUDED
Built for local experiments
One CLI. All shipped features included.
Install, inspect, compare and share. Your files and experiments stay under your control.
Choose what runs
- Without a model
- BM25 and deterministic sentence extraction need no LLM or API key.
- Local models
- FastEmbed embeddings and rerankers need the semantic extra and a first model download.
- Your providers
- Ollama and OpenAI-compatible operations use your server or API access. Remote services receive the text needed for the selected operation.
Python 3.12 · installed-package checks: Linux x64, Windows x64, macOS ARM64/Intel. Model support depends on the selected model and hardware.
06 / DOCUMENTATIONMAKE YOUR NEXT DECISION WITH EVIDENCE