patent-retriever-bigquery_skill

Overview

This skill is a Google Patent BigQuery–based patent retrieval tool that builds topic-driven query plans and returns structured JSON results for analysis. It focuses on collecting patent records (default at least 20) and validating outputs against provided schemas. It does not provide legal opinions or patentability conclusions.

How this skill works

The skill runs an initial seed retrieval from Google Patents data in BigQuery using user keywords, then generates a concept scan and a multi-step query plan. It executes the planned queries to gather patent records and outputs normalized JSON files (concept_scan, query_plan, retriever_raw, retriever_result). It also includes schema validation to ensure each patent entry contains required fields like publication_number and title.

When to use it

• You need a reproducible, data-driven prior art search across Google Patents.
• You want structured JSON output for downstream analysis, ML, or reporting.
• You need to iterate query plans based on initial seed results or concepts.
• You require BigQuery-scale access to patent metadata and full-text where available.
• You want validated outputs to integrate with other pipelines or tools.

Best practices

• Configure Google environment variables before running: GOOGLE_APPLICATION_CREDENTIALS and GOOGLE_CLOUD_PROJECT.
• Start with a broad seed retrieval (larger limit) to capture diverse concepts for building the query plan.
• Review concept_scan.json and adjust topic/keywords to refine the plan before large-scale runs.
• Set a sensible --min-results threshold (default 20) and monitor query coverage to avoid missing niche patents.
• Validate outputs with provided schemas to catch missing fields early.

Example use cases

• Prior art collection for technical landscape analysis on an AI-related topic.
• Generating a labeled seed set for patent classification or clustering experiments.
• Periodic automated pulls to keep a patent database updated for a specific technology area.
• Feeding structured patent metadata into ML pipelines for novelty detection or trend analysis.