Ranking & reranking

Once search has gathered candidate chunks, ranking decides which come back first. The relevant settings sit in the Ranking and Rerank sections of the retrieval settings dock.

​

Weights

The knob you’ll touch most often. How much to trust vector vs. keyword.

• Vector weight (default 0.60): cosine similarity. Strong on semantic search.
• Keyword weight (default 0.40): exact matching. Strong on identifiers.

​

Fusion

How the two scores are combined.

• Weighted (default): blend the scores with the weights above. Slider positions apply directly.
• RRF: combine by rank instead of raw score. Robust to score-scale differences, and favours chunks that rank consistently well in both vector and keyword.

​

Minimum similarity

A pass mark to enter the results. Each chunk gets a similarity score against the query, and anything under Minimum similarity is dropped. The default is 0, which means there is no cutoff. Even loosely related chunks come back to fill Top k. If irrelevant chunks keep sneaking in, nudge this up.

​

Rerank

The reranker re-scores search results so more relevant chunks float higher (off by default). Biggest quality lift you can get, but adds 200–500ms. Pick a provider:

• Off (default)
• Cohere: rerank model picks from rerank-v4.0 / rerank-v3.5.
• Jina: picks from jina-reranker-v2-base-multilingual / jina-reranker-v1-base-en.

Turning rerank on also reveals a Pool field that controls how many candidates the reranker receives before re-scoring. Larger pools give the reranker more to work with at the cost of a bit more latency. You’ll notice a real bump in top-result accuracy on hard queries. Turn it off wherever sub-second responses matter.

​

KG boost

When a causal graph is linked, chunks attached to KG nodes can be weighted higher. See Causal reasoning.

​

Watching it work

Every result on the retrieval page carries its similarity score. Change a setting, rerun the same query, and see how the ranking shifts. For preset-level comparison, Compare mode is faster.