Pharma
Research Patent Cliff Competitive Intelligence
In this example, we built an app that uses a Nimble Web Search Agent to search the web for medicines nearing patent expiry and return a cited report on FDA approvals, legal dockets, pricing signals, manufacturer response, and near-term patent events.
Quick Start
Inputs
- Medicine list A plain-language list of medicines and patent-expiry context entered in the browser.
- Optional JSON or CSV list A reusable medicine list passed to the CLI with DRUG_LIST_PATH; entries can include name, expiry, and notes, and the app wraps the file in the same standing research brief the browser uses.
- Research brief The default brief asks for FDA approvals, court dockets, pricing, manufacturer responses, and patent events in the next 24 months, adaptable to another therapeutic area.
- Server-side configuration NIMBLE_API_KEY and NIMBLE_AGENT_ID for the research itself, plus one model-provider key for the single Vercel AI SDK call that starts each run.
Outputs
- A Nimble research run ID returned immediately, without holding the browser request open
- An in-progress state that resumes after a browser refresh in the same session, and a finished report that survives a refresh too
- Per-medicine findings on biosimilar and generic approvals, docket activity, pricing, manufacturer response, and upcoming patent events
- A final report with source links, whether the agent is configured for prose or structured JSON output
- Nimble confidence metadata, including confidence level, cited-claim count, and source count
How it works
A 6-step asynchronous research pipeline.
- Define the research task The user enters medicines and patent-expiry context in the browser, or supplies a JSON or CSV list to the CLI; either path is wrapped in the same standing research brief.
- Start Nimble research The Next.js API route calls Vercel AI SDK generateText() pinned to Nimble's nimbleAgentStartRun() connector tool, which starts the Web Search Agent run and returns a run handle.
- Keep only the run handle in the browser The UI stores the returned runId in sessionStorage, while NIMBLE_API_KEY and the Nimble Agent ID stay server-side.
- Poll for the result The browser calls the same API route every five seconds; the server invokes nimbleAgentRunResult() directly with no model in the loop, returning the active state until the run completes.
- Render cited findings The app renders the finished markdown report with Nimble's trust metadata (confidence, cited-claim count, and source links), linking each [n] callout to its cited source via trust.claims[].callout.
- Resume without reconstructing context A refresh in the same session resumes an in-progress run or restores a finished one, and the CLI can retrieve any report later from its task_run_... handle.
Stack
Nimble primitives plus the full runtime stack.
Nimble APIs
What it does
- Web Search Agent A pre-created wsa_... agent instance performs the long-running, source-backed patent-cliff research.
- @nimble-way/ai-sdk Nimble's Vercel AI SDK connector, registering nimbleAgentStartRun(), nimbleAgentRunStatus(), and nimbleAgentRunResult() as AI SDK tools that every call to Nimble goes through.
- Nimble trust metadata Supplies the confidence level, cited-claim count, and source list rendered with a completed report.
3rd Party Tools
Role
- Vercel AI SDK v6 + OpenAI provider Runs a server-side agent that starts each research run with one forced tool call, on gpt-5-nano by default and swappable for any AI SDK provider.
- Next.js 16 App Router UI and the Node.js API route that starts and polls runs, with a 60-second ceiling since it never waits for the full research job.
- React 19 + react-markdown Renders browser input, progress state, session persistence, and the markdown report with linked citation callouts.
- TypeScript Typed server, client, and CLI implementation across the app.
Reach out if you have any questions.
Talk to an Expert