Methodology
Our design principles.
How Relevan designs search for LLM systems that need accuracy, transparency, and lower token burn.
The learning loop
Relevance is an ongoing process, not a one-time launch task. Relevan sees every search and refinement in a session on its own. You report the outcome, such as a click, an agent judgment, or a conversion, with one call to the feedback endpoint.
Capture failures and near-misses from production traffic.
Prioritize tuning by impact and recurrence.
Apply incremental ranking updates.
Measure quality deltas over time.
Core principle
LLM applications search better when retrieval is understandable. If an agent can inspect why a result matched and how to improve the next query, it can run a reliable feedback loop instead of blindly retrying.
Retrieval response model
Every Relevan search response is designed to include:
Ranked documents
Field-level evidence
Compact explanation signals
Suggested next actions for tuning
This gives the calling agent the context it needs to decide whether to accept the result, refine the query, or collect more evidence.
Schema-first planning
Before querying, agents can inspect index schema and available constraints. This reduces invalid query generation and improves first-pass recall.
Typical flow:
Inspect schema fields and data types.
Build constrained search plan.
Execute query with explanation output.
Feed interaction outcomes back into tuning.
Token efficiency model
Token burn in agentic retrieval is often driven by repeated reformulation attempts. Relevan reduces this by:
- improving first-query quality via schema-aware planning,
- exposing compact explanation signals to avoid full-context retries,
- and closing the loop so hard queries improve after each iteration.
Practical outcome
Engineering teams get retrieval that is easier to debug, safer to automate, and cheaper to operate at scale.