Five Checks to Verify Web-Grounded Answers from HostedWebSearchTool in .NET

·1 views

A developer argues that adding HostedWebSearchTool to an agent does not make it retrieval-grounded, since the tool only marks a capable backend that may browse and does not search locally. The article proposes five checks to run on the response after the marker returns, starting with confirming that search calls, result URIs, or citation annotations actually appear in the output. It recommends using the Responses API path because Chat Completions can return HTTP 400 errors with certain tool_choice settings. The author also advises enforcing domain allow and deny lists in the application process rather than relying on provider settings, failing closed when configuration is missing. Denied sources should be omitted from clickable citations, and citation records should be merged so richer annotations with dates and titles are preserved.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Security Report Separates Evidence Levels in Apex Accelerator npm Malware Case

Elastic Security Labs documented a campaign, tracked as REF8801, involving malicious npm packages @apexacc/cli and @apexacc/audit that posed as AI and Web3 tools. The reported findings include malicious launch behavior, attacker-controlled model routing, a persistent MCP server, fabricated audit results, and a macOS component called CloudSyncD alongside a staged Windows information stealer. The article argues that reports should distinguish between levels of evidence, such as package registry records, local artifacts, process and network activity, and confirmed data theft. It cautions that external observation counts do not establish the number of affected organizations. The analysis stresses that a possible credential exposure calls for different response steps than confirmed exfiltration.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Users Often Verify AI Search Answers by Clicking Through to Sources

A new study by Fractl and Search Engine Land finds that many searchers read an AI-generated answer and then click through, re-query, or seek a source to validate it. Based on a survey of 1,007 U.S. adults, clickstream data, and a query audit, the research reports that 56% of respondents click a link to double-check an AI answer, and 71% want a real source before buying. Observed zero-click rates fell from 43% in 2024 to 34% in the tracked dataset, while re-queries rose from 54% to 63%. The authors caution that these are directional observations rather than a population-wide measurement of all searches. They argue that publishers should offer evidence, expertise, and decision support that AI summaries cannot fully replace.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

CursiveLines: A Free Site for Learning Cursive One Letter at a Time

CursiveLines is a free website designed to help people learn cursive handwriting one letter at a time. Each letter shows numbered strokes and direction arrows, with playback and a trace pad for practicing by finger, mouse, or stylus. Users can also type their own text and print it as a practice sheet on A4 or US Letter paper. The site covers five cursive styles, a print course, a handwriting photo analysis tool, and a signature generator, all without requiring an account. The creators built it as a Next.js site, using hand-placed pen paths rather than a font.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Implied vs. Historical Volatility: Which Better Predicts Future Realized Volatility?

A discussion piece examines whether option-implied or historical volatility is the better forecaster of future realized volatility. It finds that implied volatility is generally the stronger standalone predictor, since it reflects market expectations and updates quickly as new information arrives. Historical volatility, computed from past returns, is simple to calculate and benefits from volatility clustering, but it looks only backward. The author notes implied volatility often overshoots subsequent realized volatility because option prices embed a volatility risk premium. The suggested practice is to combine horizon-matched implied volatility with rolling realized or GARCH-based estimates, calibrate for bias, and validate across regimes and horizons.

0 comments Read more at DEV Community