searchMode parameter controls who decides how much work each search gets: Andi, or you.
auto(default) — Andi reads each query and sets the compute, models, and search depth it needs. No tuning required.- Fixed modes — pin an effort level (
instant,low-cost,fast,balanced,deep,exhaustive) when you want the same behavior on every call.
Automatic mode
When you omitsearchMode (or set searchMode=auto), Andi decides per query how much effort the search needs. Simple lookups and navigational queries resolve fast. Complex, multi-faceted, or research questions get deeper treatment — query expansion, more search angles, follow-up searches, stronger reranking, spell correction. Multi-query requests get deep coverage automatically.
metrics.cost_dollars.
Use auto unless you have a specific reason not to. It adapts as your query mix changes, so you never have to retune mode choices as your application grows.
Setting the effort yourself
When you want the same effort on every call — a predictable latency, a cost ceiling, or research-grade coverage regardless of query — pin a mode:
These are typical response times. To target a specific response time instead, use the
timeout parameter: omit it for the mode’s normal speed, or send a value to have searchMode=auto step down to the richest mode that fits it. It’s a target, not a hard ceiling — a larger value lets a slower mode take up to that long.
Choosing a mode
-
Use
instantwhen speed matters most. It returns at most 10 results and no page content. -
Use
fastfor quick, simple lookups in user-facing search and high-volume batch jobs. -
Use
low-costwhen spend matters most and you can trade some speed for coverage. -
Use
balancedfor general-purpose search when you want predictable mid-range latency. -
Use
deepwhen result quality matters more than speed: research questions, multi-faceted topics, queries that might contain typos. -
Use
exhaustivefor questions with several parts or steps — for example comparing two things, or finding one fact and then another that depends on it.
auto never picks instant or low-cost. Speed and cost are your choices to make; auto picks the mode the question needs. With instant, a limit above 10 returns 10 results, and a request for page content or extracts returns none; the response’s warnings field says so. effort does not apply to instant.
What higher effort buys
Moving up the ladder does more than allow extra time. Deeper modes expand your query into related variations, fan out follow-up searches based on what the first pass finds, and apply stronger semantic reranking.auto uses the same machinery, scaled to what each query needs.
The effort parameter
effort pins the same ladder as searchMode, using generic tier names instead of Andi’s mode names — a second way in for agents that already speak a generic effort convention.
searchMode always wins over effort — if you send both, the named mode applies. Left at the default auto, an explicit effort pins its equivalent mode, same as naming that mode directly. Inside low-cost, effort shapes thoroughness within that lane’s own pricing and limits rather than leaving the lane.
Invalid values return a 400 listing the valid tiers: low, medium, high, max.
The resolved tier is echoed back in metrics.effort (JSON) and effort in the frontmatter (format=context):
The reranker parameter
The reranker parameter controls the strength of semantic reranking applied to your results:
Stronger reranking improves result ordering for complex or ambiguous queries, at the cost of additional latency. The
auto default picks an appropriate level based on your searchMode and query.
Pricing
Search requests are priced on outcome: the charge reflects the work the search performed, the results it returned and the text in them (see how a charge is built). Withauto, that means the price follows the effort Andi chose for each query. With a pinned mode, cost is more uniform call to call. The billed amount (the values in the examples below are examples only) is returned in every response as metrics.cost_dollars (JSON) or cost_dollars in the frontmatter (format=context).
See pricing for how every endpoint is priced and what is not charged.
MCP tool
Agents using the Andi Search MCP server control the same behavior through theandi_web_search tool — its searchMode and effort parameters accept the values on this page.
Examples
Auto mode (recommended)
Pinned fast mode for a real-time UI
Pinned exhaustive mode for research
Next steps
Fast mode
When to pin fast mode and what to expect.
Deep mode
Multi-angle search with spell correction.
Query parameters
Full parameter reference.
Response format
Response structure and metrics fields.

