Research mode: investigating across several sources
How the assistant builds a plan, runs tasks in the background, shows progress and returns a report with a source legend. How long it takes.
An ordinary assistant answer is limited to one turn and one context. When a question requires going through hundreds of records from several sources and calculating something from them, the assistant either does not read everything or calculates approximately. Research mode solves this differently: a plan of tasks, execution in the background, calculations done by code, a report with sources.
When you need it
- A market overview or a selection across hundreds of listings, records or documents.
- Reconciling data from several 1C bases, or from 1C and a knowledge base, in a single question.
- A legal brief across statutes and internal regulations at the same time.
- Checking a document against criteria from another document.
- Any calculation of medians, shares, ratings or monthly totals where you need accuracy, not an estimate.
You do not need the mode for a simple reference question: an ordinary answer with sources is faster.
How it works
- You ask the question as usual. It helps to name the sources, the period and the result format right away.
- The assistant builds a plan of tasks: which sources to read and what to calculate. Tasks can depend on each other or run in parallel.
- The assistant ends its turn and reports that the investigation has started.
- Each task runs in the background in a separate executor with a log of every step. Reading comes from the connected sources: knowledge bases, conversation documents, attachments and connectors (1C, Bitrix24, external services through MCP).
- Calculations are performed by code in an isolated sandbox over the collected data, not by the model from memory.
- When it finishes, a system message arrives in the conversation, the assistant reads the results and gives an answer: tables, conclusions, an action list and a sources section.
The data of a completed investigation remains available in the conversation. A follow-up question such as "compare with the archive by sector" runs only a calculation on top of the data already collected, without going back to the sources.
Where to follow progress
Running investigations are visible in the background tasks panel. In the "Task activity" block, each task shows its current step, tool, duration, number of retries, coverage and the time of the last update. If a task failed, the reason is visible.
If a calculation ended with an error, the investigation is marked as partial: the data from all reads is saved, the assistant sees which task did not complete and can restart only the calculation.
Note
An investigation only reads. Tasks do not edit documents and do not write to external systems. Actions that require approval are performed separately, in an ordinary dialogue.
How long it takes
The duration depends on the number of tasks, the volume of sources and the speed of the model. Reference figures from real runs:
| Scenario | Investigation time |
|---|---|
| Three reads from 1C and one from a knowledge base, then a calculation | about one minute |
| Six tasks: an archive from a knowledge base, four parallel reads of listings, a calculation | about 2.5 minutes |
| Four tasks over three documents with a check matrix | about 10 minutes |
| A follow-up calculation on top of the collected data | about 30 seconds |
On top of this come the planning turn before the start and the answer turn after the wake-up, usually 30-60 seconds each. In a typical case, 3-5 minutes pass from the question to the answer. While an investigation is running, you can keep working in other conversations.
How to read the report
In the answer, the assistant gives a source legend: the names of documents and their versions, the 1C bases and the tools from which each figure was obtained. If a rule or a fact is not found in the sources, the assistant says so directly instead of giving information from memory. Like any answer, the report should be checked against the sources before you make a decision.
How to turn it on
The mode is turned on by the assistant's creator in its settings. If the assistant answers in the ordinary way to a question that requires heavy data collection, ask your administrator whether the investigation mode is turned on and whether the selected model supports it: not all models can call the planning tool.
Tip
Phrase the question so that a plan is easy to build: list the sources, the period, the breakdown (by month, by company, by sector) and the required indicators.
Limitations
- Knowledge base search inside an investigation requires an embedding model configured for the organization.
- The "partial" state is sometimes shown for successful investigations if a task used an aggregate tool. This does not affect the answer.
- The cost of one investigation (tokens, calls, time) is limited by a budget that is set per assistant.
Related articles: Chat with sources, Documents and reports in chat.
Ask Sema