Generators 

Two AI cost paths 

nr-llm guards completion calls with its budget middleware automatically (via the beUserUid on ChatOptions). Its specialized services (TTS, image) are not middleware-guarded, so before spending on them each generator calls AbstractGenerator::specializedAllowed(), which checks nr-llm's BudgetService and the service's isAvailable(). A budget-starved run therefore still yields the cost-free variants (for example the deterministic HTML Schaubild) while skipping the AI-image variants.

Cost attribution 

Every nr-llm call names this extension and the pipeline step it belongs to (AbstractOptions::withCallerSource()), so nr-llm's Analytics module breaks usage and cost down by nr_repurpose and by operation instead of listing them as "Unattributed". The operation names are constants on NetresearchNrRepurposeServiceCallerSource:

analyzeDocument        document synthesis (and its corrective retry)
analyzeDocumentChunk   map step, one call per chunk of a large document
extractPdfVision       OCR of one rasterized PDF page
generatePodcast        podcast dialogue script
generateDiagram        Schaubild diagram body
generateStory          story carousel copy
generateExecSummary    executive summary
generateFaq            FAQ question/answer pairs
generateSocialPost     social posts, all platform variants in one call
generateNewsletter     newsletter text
generateSlideDeck      slide deck titles and bullet points
generateHandout        handout title, lead, sections and key facts

ConfiguredCompletionService, the decorator every text completion passes through, stamps the extension key on options that carry none, so a new call site is attributed even if it forgets. The operation stays with the call site — only it knows which step it is.

The specialized calls (TTS, image generation) carry no attribution: nr-llm's specialized services do not read the caller source. The PDF vision annotation is set but does not reach the telemetry row either, because VisionService::analyzeImageFull() rebuilds the options object without it (netresearch/t3x-nr-llm#845).

Podcast 

PodcastGenerator asks the completion service for a dialogue script sized to the document scope, spoken by the job's selected personas (one to three, each with its name, character description and optional own TTS voice) or by the default hosts (Host A = nova, Host B = onyx). Each turn is one specialized TTS call producing an MP3 segment, with a single retry on a transient failure and a skip (rather than a whole-episode failure) if a turn still fails. The segments are concatenated by FfmpegAudioStitcher::concat() (ffmpeg concat demuxer, stream copy, no re-encode); per-segment durations are read with ffprobe and fed to WebVttBuilder so the subtitle cue times match the audio. The MP3 and the .vtt are stored in FAL; the speaker-tagged transcript is kept on the artifact row.

Schaubild 

SchaubildGenerator produces three artifact rows for empirical comparison:

• html — the LLM writes a branded HTML diagram body; it is wrapped in the theme template and rendered opaque to PNG by Chromium. No specialized call, so this variant always proceeds.
• html_bg — an AI background image plus the same diagram rendered transparent, composited together (see Rendering toolchain).
• ki_image — a full AI text-to-image from a content-derived prompt.

The diagram is rendered at 1200 px wide, auto-height.

Instagram story 

StoryGenerator asks the completion service once for the whole carousel — a cover slide, one slide per key point (at most four) and an outro with the source attribution, capped at six slides; the planned cost scales with the expected slide count. Each slide is rendered from the branded 9:16 template (1080×1920) into its own artifact row (variant slide-N; slide role, index and total in the metadata), so a failed slide render fails only that slide. When the image service is available and within budget one portrait AI background is generated and composited behind every slide; otherwise the slides fall back to flat renders.

Text formats 

ExecutiveSummaryGenerator, FaqGenerator, SocialPostGenerator and NewsletterGenerator extend AbstractTextGenerator. Each makes one CompletionServiceInterface::completeStructured() call with its own JSON schema; nr-llm validates the answer against the schema and asks once more with the validation failure when it does not match. The generator then applies what a schema cannot express — caps, the platform character limits (TextLimiter, cut at a sentence boundary, or at a word boundary when that would keep less than half), hashtag normalisation — and stores the plain text in script_text and the structured answer in metadata.content. The format's task sits in the system prompt; the user prompt carries only the source-derived brief as untrusted data inside <source_material> tags, with tag-like <source… sequences in the data neutralised. The social posts become one row per platform variant (linkedin, x, instagram); the other formats one row each. No file is written to FAL. See ADR-004: Text Formats as Schema-Validated Structured Output.