Kenning
whale·road
03 · Editor · retrieval-01-tune
← back to Under the Hood

Tune the pipeline

core

Pipeline · your configuration

retrieval
top_k
chunking
chunk_size
overlap
child_size
auto_merge_threshold

every change re-runs the corpus

Graded on

Evidence arrives intactRequired×4.0
No wrong-entity evidenceRequired×3.0
Context stays cheapRequired×2.0
Best evidence ranked first×1.5
Required — a miss caps the score, whatever the rest do.

Brief

Below is a working pipeline configuration. It is also the naive one: fixed chunks, pure vector search, no metadata filter. Submitted as-is it answers a little over half the questions, and four fifths of what it delivers comes from the wrong company or the wrong year.

Your configuration will actually be run against the lab corpus. The score is the measured result, not an opinion about your choices.

Task

Edit this configuration so that the evidence for at least 95% of questions arrives intact, nothing from the wrong entity or period is delivered, and the average context stays under 750 tokens.

Available settings — chunking (fixed, recursive, parent_child), chunk_size, overlap, child_size, table_merge, metadata_filter, retrieval (sparse, dense, hybrid), top_k, rerank, resolve_parents, auto_merge, auto_merge_threshold.

Running the corpus …

Running the corpus …
your configuration is run and measured — resubmit freely