> ## Documentation Index > Fetch the complete documentation index at: https://docs.topk.io/llms.txt > Use this file to discover all available pages before exploring further. # Research > Perform deep multi-step research across your document corpus. Research turns complex queries into a series of targeted searches, gathers evidence across documents, and produces a grounded answer with source citations. ## How it works When you run a **Research**, TopK: Research interprets the query in context, including the evidence gathered so far. It decomposes the problem into focused sub-queries. It runs targeted searches across your data to find the most relevant information. It pulls out concrete, source-backed facts from each result, preserving citations. It combines the gathered evidence, performing comparisons or aggregations when needed, to produce a single, grounded answer with citations. ```mermaid theme={null} flowchart LR Q[Research query] S1["Search (Step 1)"] S2["Search (Step 2)"] S3["Search (Step 3)"] E[Evidence] A[Grounded answer] Q --> S1 Q --> S2 Q --> S3 S1 --> E S2 --> E S3 --> E E --> A ``` ## Usage Research is invoked through `ask` by setting `mode` to `research`. ```bash theme={null} topk ask "What was the 2022 and 2023 total value of commercial loans modified as TDRs, and the total 90+ days past due for consumer loans at December 31, 2024?" -d finance --mode research ``` ```python Sync theme={null} import os from topk_sdk import Client client = Client( api_key=os.environ.get("TOPK_API_KEY"), region="aws-us-east-1-elastica", ) for message in client.ask( "What was the 2022 and 2023 total value of commercial loans modified as TDRs, and the total 90+ days past due for consumer loans at December 31, 2024?", ["finance"], mode="research", ): print(message) ``` ```python Async theme={null} import os from topk_sdk import AsyncClient client = AsyncClient( api_key=os.environ.get("TOPK_API_KEY"), region="aws-us-east-1-elastica", ) async for message in client.ask( "What was the 2022 and 2023 total value of commercial loans modified as TDRs, and the total 90+ days past due for consumer loans at December 31, 2024?", ["finance"], mode="research", ): print(message) ``` ```typescript theme={null} import { Client } from "topk-js"; const client = new Client({ apiKey: process.env.TOPK_API_KEY, region: "aws-us-east-1-elastica", }); for await (const message of client.ask( "What was the 2022 and 2023 total value of commercial loans modified as TDRs, and the total 90+ days past due for consumer loans at December 31, 2024?", ["finance"], undefined, "research", )) { console.log(message); } ``` Here's an example of a **Research** query over a financial dataset: **Query:** What was the 2022 and 2023 total value of commercial loans modified as TDRs, and the total 90+ days past due for consumer loans at December 31, 2024? **Answer:** * Based on exact phrasing matches to corporate filings, the question refers to Bank of America. During 2022, the carrying value of Bank of America's commercial loans modified as Troubled Debt Restructurings (TDRs) was \$1.9 billion. 1 * There is no 2023 value reported, as the TDR reporting designation and accounting guidance were eliminated effective January 1, 2023, upon the adoption of ASU 2022-02. 2 * As of December 31, 2024, Bank of America's total consumer loans and leases that were 90 days or more past due totaled \$2,422 million. 3 **References:** ## Scoping the search For guidance on: * querying across specific datasets * applying document filters * limiting which documents can be considered during retrieval see [Scoping the search](/core/ask#scoping-the-search) in the Ask guide. ## When to use Research Use Research for analytical questions that require more than a single retrieval pass to answer reliably. It is intended for multi-hop questions when searching, connecting, comparing, and synthesizing facts is needed to answer the question. Use **Research** when: * the answer depends on multiple facts from different parts of a document or different documents * the query requires comparison across years, entities, filings, or sections * the answer requires complex computation, aggregation, synthesis or resolving intermediate facts * you care more about depth and correctness than getting the fastest possible answer Prefer [Ask](/core/ask) with `summarize` mode when: * the question is direct and likely answered by one or a few passages * you want the fastest grounded answer * the extra decomposition and multi-step reasoning are not necessary to reliably answer the question Because **Research** performs more work than an `ask` query in `summarize` mode, it will usually take longer to produce an answer.