Frequently Asked Questions

GovQuery

What is GovQuery?

GovQuery is an AI-powered search engine that makes key government oversight sources easier to explore. It searches across multiple federal document repositories — and within the documents themselves — linking results to the exact page where a passage appears.

What sources and how much data does it cover?

GovQuery indexes 309,998 documents and articles, as well as 18,897 open oversight recommendations across these sources:

How far back do the data and reports go?

GovQuery captures all open recommendations from Oversight.gov and GAO's public recommendations database, GAO reports published since 2010, OIG reports since 2000, and any earlier reports referenced in oversight recommendations. Note that roughly 15% of the reports listed on Oversight.gov do not have a working link, typically when referencing external agency websites.

How does GovQuery differ from the search on Oversight.gov and gao.gov?

Those platforms excel at finding a report by its description and typically return one link per report. GovQuery searches inside reports and links to the exact page, adds AI-assisted image search, AI-generated tags, and filtering, and searches across multiple sources at once.

What does GovQuery offer that a web search engine like Google does not?

Public search engines surface results at a high level — usually one link per report. GovQuery links directly to the relevant individual pages and lets you run targeted searches of open recommendations across oversight bodies.

How does the agency filter work?

The filter narrows results by associated agency. GAO agencies come from GAO's official filter; Oversight.gov uses its "Agencies Reviewed/Investigated" field; CRS reports and executive orders use an AI prediction based on the agencies a document mentions. Parent agencies are included when a sub-agency is identified, and names are mapped to USAspending.gov agency codes.

Recommendation Spotlight

What is the Recommendation Spotlight?

The Recommendation Spotlight is a curated, searchable tool for federal oversight recommendations. It draws from GAO reports and recommendations and from federal OIGs via Oversight.gov, combined in PIA's custom data lakehouse. Recommendation text appears exactly as published — we don't rewrite it.

What is its purpose?

It helps policymakers, civil servants, and researchers identify evidence-based findings to improve government and safeguard integrity. By organizing oversight data into one place, it supports targeted searches, thematic analysis, and comparison across agencies.

What makes it different from other tools or databases?

Unlike generic search engines or individual agency websites, the Spotlight combines recommendations across multiple oversight platforms into one standardized interface, enriched with structured metadata, AI-assisted classifications, and filters that support deeper analysis.

Data sources and content

What are the specific data sources used to build the Spotlight?

• GAO reports and testimonies and open oversight recommendations. GAO provides non-partisan information that can be used to improve government.
• Federal OIG reports and open recommendations via Oversight.gov (consolidated from members of the Council of the Inspectors General on Integrity and Efficiency). Federal OIGs work to prevent and detect waste, fraud, and abuse in their agency's programs and to promote economy, efficiency, and effectiveness.

Does the Spotlight include every recommendation from GAO and federal OIGs?

It covers close to 100% of the recommendations made publicly available by GAO and Oversight.gov. Some recommendations are not publicly released or are missing key information, which can affect aggregated and filtered views. We do not have direct access to internal GAO/OIG systems and rely on public sources, so we encourage cross-checking against the GAO or individual OIG websites.

How often is the Spotlight updated?

Once a week, on weekends, to avoid peak hours. When a new recommendation appears depends on the source's own release and our processing timelines.

Do you edit or adapt the recommendations?

No. Recommendations appear exactly as sourced, with no edits. AI-assisted fields (themes, tags, integrity-related prediction) are added for analysis and rely on the original recommendation text.

Does the Spotlight include state and local oversight bodies?

Not currently — it includes only GAO and federal OIGs reported on Oversight.gov, though these recommendations often concern federal spending that flows to other levels of government.

Fields in the recommendations data

The recommendations dataset includes the following fields:

• Source — GAO, or a federal OIG via Oversight.gov.
• Entity — the reviewed entity (top-tier or sub-tier agency). GAO sets this to "Congress" for Congressional matters.
• Top-Tier Agency — the federal agency per USAspending.gov agency codes (unmapped: under 0.5%).
• Recommendation — the specific recommendation from the report.
• Report — the title of the report containing the recommendation.
• Report hosting page — the GAO or Oversight.gov publication page.
• Report URL — a direct link to the downloadable PDF.
• Status — open or resolved (the Spotlight emphasizes unresolved recommendations).
• Age — how long the recommendation has remained open (grouped by years).
• Integrity-related (PIA AI) — a model prediction of whether the recommendation connects to program integrity; "No" if uncertain.
• Priority flag — GAO's "Priority" column or Oversight.gov's "Significant Recommendation" designation.
• GAO Audit Topics — GAO-assigned subject areas.
• Fraud Risk Management Theme (PIA AI) — predicted category: Governance & Capacity, Risk Assessment, Design & Implementation of Controls, Detection, Response & Evaluation, Data & Technology, or Other (based on GAO's Framework for Managing Fraud Risks and the CFO Council Program Integrity Playbook).
• Matter for Congress (PIA) — flags Congressional action; the agency is set to "Congress".
• Report Type — the Oversight.gov report type, or "Report" by default for GAO.
• Recommendation URL — the Oversight.gov recommendation link, or the report hosting page.
• Issue date — the official release date.
• Federal Fiscal Year — the government fiscal year (October 1 – September 30).
• Covid Year — whether issued during 2020, 2021, or 2022.
• Agency Comments — the audited agency's responses (GAO only; Oversight.gov pending).
• Report Author — the report author (GAO or the auditing OIG).
• Report Issue Date — the full report publication date.
• Report ID — the unique GAO/Oversight.gov system identifier.
• Data Updated (PIA) — the most recent date the record was added or updated.
• PIA Id — an internal PIA tracking number.

Known limitations and caveats

We aim for accuracy and transparency. The following known issues affect the data:

Oversight.gov recommendations

• Our totals match the GAO website; minor discrepancies with Oversight.gov exist because that site is complex to parse. About 1.5% (≈280) more recommendations appear on Oversight.gov than we extract.
• Priority flags can be inconsistent on the source: as of August 7, 2025 the Oversight.gov priority filter returned 66 recommendations, while individual pages showed roughly 3,500 marked "Significant recommendation: Yes."
• About 15% of recommendations lack a corresponding "View Report" or PDF link, and some are missing source-system fields.
• About 2% have blank recommendation text, and about 2% have text redacted for sensitivity or security.
• Seven recommendations lack a report/agency association, and fewer than 0.5% are duplicated in the source system.

GAO recommendations

• "Matters for Congress" are assigned to the agency "Congress" even though they relate to specific agencies.
• Sensitive reports are redacted, and their associated recommendations are likely excluded.
• About 0.3% of recommendations don't map to USAspending.gov agency data because of name misalignment.

Technology and use of AI

How does PIA use AI / large language models?

LLMs extract, summarize, and classify recommendations from lengthy oversight reports, identify key themes, and assign tags that improve searchability — particularly useful where there isn't enough labeled data for traditional machine learning. AI also assists our data-processing pipelines and software development.

How do you ensure the accuracy and reliability of AI outputs?

All AI outputs are grounded exclusively in GAO and Oversight.gov source data — never in a model's raw web training data. Outputs are evaluated automatically and reviewed by subject-matter experts before inclusion. Automation combined with human oversight maintains quality and trustworthiness.

How do you develop the LLM process and test its accuracy?

We write expert-defined prompt rules, generate initial predictions to reveal the distribution, sample a balanced 200-record evaluation set for human verification and correction, re-run the model and compare to the human-reviewed labels, calculate accuracy/recall metrics, analyze mismatches, and refine the prompt and dataset. We repeat the cycle until results stabilize and apply advanced techniques (e.g. chain-of-thought prompting, different models/settings). A final evaluation runs automatically in the data pipeline to track accuracy drift.

Which LLMs or AI tools are used?

Only Microsoft Azure–hosted models, so data is not shared beyond PIA. GPT-4.1 and GPT-4.1 mini are used for recommendation tagging, AI search (indexing documents and summarizing results), and software-development support. All calls include real-time granular logging and content-safety filters.

How do you extract and classify recommendations from reports?

We combine natural-language processing, prompt engineering, and rules-based filtering to extract recommendation text and assign categories from defined taxonomies. Human review ensures consistency.

Is the AI summarizing recommendations or just helping retrieve them?

Tabular recommendations are presented as-is, without summarization. AI is used in Search to provide concise result summaries with supporting citations.

How do you prevent bias or errors introduced by AI?

Grounding AI responses exclusively in GAO and Oversight.gov data mitigates the bias risks associated with open web content, and human review assesses outputs for accuracy and bias.

How are you using AI for software development?

AI accelerates development of our data pipelines, user interface, and website assets via the Cursor AI code editor and GitHub Copilot.

Verification and trustworthiness

How do you verify recommendations are accurate and up to date?

All recommendations are linked to primary source documents and reviewed for accuracy. We make updates when oversight bodies publish progress or changes about the implementation of recommendations.

Are recommendations linked to the original oversight reports?

Yes. Each entry links to its source documents, typically hosted on Oversight.gov, gao.gov, or agency websites.

Citing the data

Please cite both the original source agencies (GAO, Oversight.gov) and the Program Integrity Alliance, and link to the original recommendation or report where possible. PIA adds value through data extraction and parsing, cleaning and normalization, enrichment with metadata and classifications, and curated summaries, tools, and visualizations. Suggested form:

U.S. Government Accountability Office; various Offices of Inspector General. Data retrieved via Program Integrity Alliance (PIA), Spotlight on Oversight Recommendations, programintegrity.org.

Bookmarking

Can I bookmark searches, table settings, and trace diagrams?

Yes. The web address captures your search state and customized views, so you can bookmark or share it to return to previous work.

Feedback and future development

How can I provide feedback or request new features?

You can share feedback right inside GovQuery:

• Feedback button — the speech-bubble icon in the bottom-right corner of every page opens a short feedback form for general comments, bug reports, and feature requests.
• Rate a search — use the rating control on the AI Overview / search results to tell us whether a result was useful.
• Rate a chat answer — each AI Mode chat response has a rating control so you can flag a helpful or unhelpful answer.

We review this feedback to prioritise improvements to functionality, usability, and coverage.

How do you plan to improve the Spotlight?

We continue to improve based on user and community feedback — including refining AI-assisted fields that help auditors and researchers quickly assess relevant recommendations, and improving how we ingest and categorize newly issued oversight recommendations.