Go-To-Market
Run company research tasks with LangChain
The langchain-nimble toolkit adds Nimble Search and Web Search Agents to a LangChain agent, and the model picks the right one for each field: a fast search for a website or a head office, a cited research run for anything needing multiple sources. Column names from the input file become the run's output schema, so results map straight back to the cells they fill. Values that cannot be confirmed from a retrieved source are left empty rather than guessed.
Quick Start
Inputs
- A CSV with gaps A header row, one column holding the company name, and empty cells wherever values are needed. A CRM export works without changes.
- Governed columns Which fields must come from a cited Web Search Agent run rather than a fast search. Defaults to employee count, funding stage and total raised.
Outputs
- The original file returned filled, with column names unchanged
- A source URL for every filled cell
- A confidence grade per cell, taken from the Web Search Agent run's trust data
- Which Nimble tool answered each cell: supplied in the input, fast search, or a research run
- Empty cells where nothing could be confirmed, rather than a guess
How it works
A 6-phase pipeline.
- Wire the toolkit Construct NimbleToolkit with include_web_search_agents set to true, then hand get_tools() to create_agent. Nimble Search and the Web Search Agent tools are now available to the model.
- Skip what is known Cells that already hold a value are never re-fetched, so nothing is spent on data the file already has.
- Let the model choose For each remaining gap the agent either calls nimble_search for a fact confirmable in one lookup, such as a website or a head office, or starts a Web Search Agent run for anything needing multi-source research.
- Collect the runs A run takes minutes, so runs are started and collected asynchronously. The CSV header becomes the run's output schema, so column names pass through unchanged.
- Map trust to cells Each run returns per-claim citations and confidence keyed to a JSON path, which the app maps back to the exact cell the claim fills.
- Refuse rather than guess Any value that cannot be confirmed from a retrieved source is left empty and flagged for review.
Stack
Nimble primitives plus the full runtime stack.
Nimble APIs
What it does
- Web Search Agent Exposed to the LangChain agent as three tools: start a run, poll its status, fetch the result. Each run returns a schema-conforming answer with per-claim citations and a confidence grade.
- search Exposed as nimble_search for facts confirmable in a single lookup. The model decides which of the two a given gap deserves.
3rd Party Tools
Role
- langchain-nimble NimbleToolkit, the single entry point. Setting include_web_search_agents to true adds the Web Search Agent tools to the toolset.
- langchain create_agent handles the tool-calling loop and records which tool the model picked.
- langchain-anthropic The routing model.
- python 3.10+ Required by langchain-nimble 4.x.
Reach out if you have any questions.
Talk to an Expert