# Ranking Tuning

# Ranking Tuning

Ranking tuning is the process of making “the right results appear first” for your users. In Curiosity Workspace, tuning typically happens through configuration of:

• indexed fields and boosts
• filters and facets
• hybrid retrieval strategy (keyword + semantic)

# Start with a relevance baseline

Before changing knobs:

• define 20–50 representative queries
• capture expected “good results” (gold set)
• include edge cases (acronyms, short queries, long queries)

# High-leverage tuning levers

# Field selection

• Index only fields that should participate in retrieval.
• Remove fields that inject noise (e.g., boilerplate text).

# Field boosts

• Boost high-signal fields (titles, summaries, identifiers).
• Reduce boost for long body fields if they dominate rankings.

# Facets and scoping

• Add facets users actually use (status, type, owner).
• Use graph-related facets for meaningful constraints.

# Recency and freshness

• For time-sensitive domains, prefer a sort mode that rewards recent items.
• Consider separating “recent first” views from “relevance first” views.

# Vector tuning

For semantic retrieval:

• choose which fields are embedded (usually long text)
• tune similarity cutoffs (what qualifies as “related”?)
• tune chunking (for long text fields)

# Common pitfalls

• Tuning without evaluation: always compare before/after on a query set.
• Boosting everything: boosts should express a clear priority order.
• Ignoring filters: the best ranking often comes from better scoping, not just scoring.

# Next steps

• Implement hybrid retrieval: Hybrid Search
• Turn extracted entities into facets: NLP → Entity Extraction