Experimental tutorial · P3
Contextual retrieval,
measured honestly.
p3_contextual_retrieval is a direct P0 candidate that changes only server-owned chunk contextualization.
Auditable contract
P0 → P3Single-variable experiment
Strict-fail context generation
What changes
Only contextual text.
P3 uses a fixed server prompt and bounded inputs to situate each chunk within its verified document. Clients cannot supply prompts, models, limits, or context content.
What is measured
Index facts, not claims.
Runs persist whether contextualization was enabled, the count of non-empty contextualized chunks, and their average deterministic text length.
When comparison is valid
Only direct P0 → P3.
The API compares only a completed compatible P0 parent and a P3 run with actual contextual facts and standard evaluation records.
What it does not prove
No automatic superiority.
Contextual index metrics are not a quality, cost, latency, or confidence result. Evaluate those separately with the declared dataset.
Reproduce it
Run the controlled walkthrough.
The real browser regression starts PostgreSQL, Qdrant, migrations, API, Console, and a controlled immutable Pack. It runs P0 then P3 and verifies the returned contextual index facts.
orag — tutorial P3
$ make console-real-tutorial-clone-e2e ✓ P0 baseline completed ✓ P3 contextualization applied ✓ contextualized_chunk_count recorded ✓ direct P0 → P3 comparison ready