# Import an existing retriever

Evaluate your system’s returned spans against the same original sources.

CLI version: 0.2.1 · [Read online](<https://ragvidence.medo-id.com/docs/import-retriever>)

## Exact-span imports

**results.jsonl · schema illustration**

```
{"question_id":"Q1","hits":[{"source_id":"COPY_ORIGINAL_SOURCE_ID","start":0,"end":13,"text":"eight seconds","rank":1,"score":0.8}]}
```

Use original source IDs and exact offsets/text from your corpus; the illustration assumes the source starts with eight seconds. Include a row for every question, even empty hit lists. Ranks are one-based; scores can be omitted. All hit tails are validated before selection.

**Terminal**

```
ragvi import-data ./docs --reviewed ./reviewed.json --results ./results.jsonl
```

## Registered-unit proxies

ID-only hits use source\_unit\_id referring to a registered original or passage. They are explicitly labeled registered\_unit\_proxy: the actual external context was not supplied. Exact spans and proxy hits cannot mix in one run. Proxy scores cannot satisfy actual-context quality criteria.

## Use your system as a comparison baseline

**Terminal**

```
ragvi compare-data ./docs --reviewed ./reviewed.json --results ./results.jsonl --techniques bm25,parent-child
```

existing becomes the default baseline; --baseline chooses another named participant. Missing external timing, internal stages and generation-request context remain unavailable.

- [Import contract and evidence validation](<https://github.com/Medo-ID/ragvidence/blob/v0.2.1/docs/portable-techniques.md>)