Airgentic Help
Limited trial — Multi-step agent retrieval is currently available for selected services only. Contact Airgentic before enabling or changing this capability.
Multi-step agent retrieval helps an AI agent answer questions that need more than one search step.
For example, if a user asks:
What events are on during Family History Month?
The agent may first need to find out when Family History Month is, then use that date range to search for relevant events. This is different from a simple search, where the agent searches once and answers from the results.
Most questions are answered with the default search-and-answer flow:
Multi-step retrieval adds a planning layer for complex questions:
This is useful when a question contains an implied condition, time period, or dependency that is not directly stated in the question.
Multi-step retrieval is most useful for questions like:
In these examples, the agent may need to discover a timeframe, policy date, eligibility condition, or related fact before it can search properly.
Default search is usually enough for:
Multi-step retrieval should not be used just because an answer is important. Use it when the agent needs to work something out before searching for the final answer.
During the limited trial, Airgentic will usually enable multi-step retrieval for you.
At a high level, the agent needs:
This is normally configured by Airgentic support or an authorised technical administrator. Most content owners do not need to edit these settings directly.
When multi-step retrieval is enabled, avoid forcing the search function to run first.
In the agent editor, the Force call first option can be useful for simple search agents, because it makes the agent search before answering. However, for multi-step retrieval it can prevent the agent from planning first.
If search is forced first, the agent may skip the planning step and search too broadly. For example, it may search for “Family History Month events” before first finding out when Family History Month is.
For multi-step retrieval, leave search function selection automatic unless Airgentic support advises otherwise.
What events are on during Family History Month?
This may still produce a plausible answer, but it is less reliable because the search was not properly constrained.
Use Admin Chat to test multi-step retrieval before making changes live.
Open the Trace Log while testing. For a good multi-step retrieval flow, look for:
| Trace event | What it means |
|---|---|
| Long-term objective | The user’s overall question was captured. |
| Retrieval plan updated | The agent created or revised a multi-step plan. |
| Get Relevant Information | The agent searched the knowledge base. |
| Select Best Search Results | The agent reviewed candidate search results. |
| Review Retrieved Evidence | The selected evidence is being used as intermediate context, not final answer evidence. |
| Fact recorded | The agent recorded a discovered fact, such as a date range. |
| Step advanced | The agent moved to the next retrieval step. |
| Answer Question | The agent has enough final evidence to answer. |
Not every question should show all of these events. Simple questions may still use a single search and answer directly.
Use a mix of simple and complex questions.
These should not over-plan:
These should usually produce a retrieval plan:
These help reveal whether the agent is guessing:
For named periods, check that the agent establishes the period before searching for final answers.
| Symptom | What to check |
|---|---|
| The agent searches immediately and skips planning | Check whether Force call first is enabled for search. Disable it for multi-step retrieval. |
| The trace does not show Retrieval plan updated | The agent may have treated the question as simple, or the multi-step retrieval tools may not be enabled for that agent. |
| The agent answers after one broad search | Check whether it created a multi-step plan. Without a plan, simple one-step behaviour is allowed. |
| The agent selects event results before confirming a named period | Strengthen prompt instructions or add clearer source content that defines the period. |
| The agent cannot find the prerequisite fact | Check whether the relevant content is indexed and searchable. |
| The answer is plausible but not well constrained | Review the Trace Log to see whether the second search used the discovered date or condition. |
| The agent keeps searching repeatedly | Check whether the query is too broad or whether content is missing from the knowledge base. |
Use Curated Answers when you need exact wording or want to pin a specific source for a known question.
Use multi-step retrieval when the agent needs to reason through a small retrieval process before answering.
Multi-step retrieval is configured per agent. It can be enabled for the Frontline agent or a specialist agent, depending on where the behaviour is needed.
Site Search is the search interface on your website. Multi-step retrieval is part of the AI agent’s answer-generation process. They use the same indexed content but serve different user experiences.
Multi-step retrieval is currently intended for text chat. Voice mode may use a different retrieval path.
During the trial:
The goal of the trial is to confirm that multi-step retrieval improves answer quality for complex questions without making simple questions slower or more complicated.