Loading today’s stories
Build1 distinct publisher2 min readPublished
Whisper into a blob container gives an agent quotable meetings with no speaker and no offset. Microsoft's answer arrives with two ingestion paths and one honest caveat about video.
The Engineer · Build desk
Compiled by The EngineerSomething wrong?How this is made
A usable citation carries fields: who spoke, where in the file they spoke, and what was on the screen while they did [4]. Transcription returns the words and drops the rest, which is why the Whisper-into-blob route yields an index that can quote a meeting without placing it [3][4]. The failure lands on retrieval rather than generation. When the question is what was decided about the vendor migration, the discriminating signal is the speaker and the moment, and neither one is in the chunk [2][4].
Content Understanding's argument is that the extractor should emit that structure during ingestion, pairing Document Intelligence's layout work with LLM reasoning over the content [5], and since Build 2026 it can run inside Foundry IQ's standard mode instead of beside it [6].
That leaves the choice of door. Path A is a property: `contentExtractionMode` set to `standard` on a Blob, SharePoint or OneLake knowledge source, no orchestration code, shipped with the Foundry IQ 2026-05-01-preview release [7]. Path B is your own analyzer run, landing Markdown and extracted fields in blob storage for an ordinary knowledge source to index [8]. The caveat in the dev.to walkthrough is worth more than the tutorial around it: Microsoft's Foundry IQ extraction announcements lead with layout, tables, figures and document-embedded images, and it is Content Understanding's own documentation that names audio and video, so whether a blob source accepts an .mp4 in your region at your API version is something to test [9]. The author's advice is to try the flag first and fall back if it does not cover your media [14].
The version arithmetic makes that fallback less optional than it sounds. The extractor is GA at 2025-11-01 [11]. The retrieval side sits at 2026-05-01-preview for exactly the extraction features you came for [12]. Combine the two and the pipeline runs at preview support, because the weakest component sets the floor [13]. Path B at least keeps that preview surface out of your ingestion contract, since the artifact on disk is Markdown and fields, which an index will still accept when the flag's API version turns over.
Older code carries its own bill. The 2024-12-01-preview and 2025-05-01-preview APIs were slated to retire on July 15, 2026, and that date has passed [11].
Field quality, meanwhile, is a deployment decision rather than a service setting. Analyzers run on models you deploy in Foundry, and the write-up reports GPT-5.2 doing better on custom field extraction across mixed layouts, domain-specific language and multilingual content, with GPT-4.1 analyzers continuing unchanged [10]. If speaker labels and slide text are the fields you need indexed, that deployment is where their accuracy comes from.
Ranked by verification strength, evidence, and original report placement.
Organizations hold years of recorded meetings, support calls, training sessions, conference talks and screen recordings, and almost none of it is retrievable.
The example given is the question "what did we decide about the vendor migration", whose answer sits inside a 47-minute recording nobody will scrub through.
The reflex fix is to bolt a transcription service onto a RAG pipeline: run Whisper, dump the text into a blob container, index it.
Such a transcript has no speakers, no timestamps that can be linked back to, no slide content and no distinction between a presenter reading a bullet and someone in the room disagreeing with it; the agent can quote the meeting but cannot say when it happened or who said it.
Azure Content Understanding ingests documents, audio, images and video and extracts the most critical information to power well-grounded generative and agentic solutions, combining Document Intelligence's traditional AI with LLM-based content reasoning.
As of Build 2026, Content Understanding is integrated with Foundry IQ standard mode for built-in content extraction inside Microsoft's retrieval and agent workflows.
Follow any of these and your For You feed starts watching them — no settings page required.
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
Single practitioner account, no primary citations
Everything rests on one dev.to tutorial. Its mechanics (property names, analyzer routing, install line) are specific and internally consistent, which is a point in favor, but no Microsoft documentation, release note or second publisher is supplied to corroborate the Build 2026 integration, the API version numbers or the retirement date. The author's own hedge on video ingestion caps how much can be treated as established.
Release milestones only, no deployments
There are dated shipping signals — a GA Content Understanding API, a Foundry IQ preview release carrying the native extraction flag, a Build 2026 integration and a preview API retirement — but no deployment, usage, customer or benchmark evidence anywhere in the cluster. Adoption is therefore scored on vendor-release existence alone.
Mildly overstated, but self-caveated
The framing — Microsoft's answer to unindexed media, two ingestion paths — runs slightly ahead of what is demonstrated, since no example of end-to-end video ingestion into a Foundry IQ knowledge source is shown and audio/video support in that path is unconfirmed. The gap stays small because the author volunteers the caveat, flags the GA/preview asymmetry, and tells readers to verify with a five-minute test rather than assume.
Practitioner tutorial with self-referential promotion
The piece is a vendor-stack tutorial on a developer platform that cross-promotes the author's earlier Foundry IQ plus LangGraph article, giving a mild attention and authority incentive toward presenting the Microsoft path as the solution. No sponsorship, affiliation or vendor relationship is disclosed in the supplied material, and the inclusion of an unflattering caveat and a passed retirement date cuts against a purely promotional read, so incentive pressure is moderate rather than high.
Low: one source over shifting preview surfaces
Low confidence follows from single-source evidence, no independent corroboration, and subject matter that is explicitly version- and region-dependent on a preview API. The durable parts — that flattened transcripts lose citation anchors, and that a GA plus preview pipeline inherits preview support — are more reliable than any specific claim about what a knowledge source will accept today.
invest
A 4B model edged the GPT-5 family at bargaining. The interesting number is the spread1 distinct publisher
build
Foundry IQ knowledge bases ship as MCP servers, and four behaviours break naive clients1 distinct publisher
product
Rillet's $100M reads as proof mid-market ERP is rip-and-replace, mostly at the cheap end1 distinct publisher
build
A UDP packet is now enough: IKEEXT RCE moves from patch queue to fire drill1 distinct publisher
Distinct publishers with included, body-backed reporting in this cluster.
dev.to
1 article · August 24, 2026