@mux/ai 0.37.0 → 0.39.0

package/dist/workflows/index.d.ts

@@ -1,4 +1,4 @@
-export { A as ASK_QUESTIONS_DEFAULT_MAX_ANSWER_OPTION_LENGTH, j as ASK_QUESTIONS_DEFAULT_MAX_FREE_FORM_ANSWER_LENGTH, k as ASK_QUESTIONS_MAX_QUESTION_LENGTH, l as AskQuestionsOptions, m as AskQuestionsResult, n as AskQuestionsType, o as AudioTranslationOptions, p as AudioTranslationResult, q as AutoCensorProfanityOptions, B as BurnedInCaptionsAnalysis, r as BurnedInCaptionsOptions, s as BurnedInCaptionsPromptOverrides, u as BurnedInCaptionsPromptSections, v as BurnedInCaptionsResult, C as CaptionReplacement, w as CensorMode, x as Chapter, y as ChapterSystemPromptSections, z as ChaptersOptions, D as ChaptersPromptOverrides, E as ChaptersPromptSections, F as ChaptersResult, G as ChaptersType, H as DEFAULT_DESCRIPTION_LENGTH, J as DEFAULT_SUMMARY_KEYWORD_LIMIT, K as DEFAULT_TITLE_LENGTH, M as EditCaptionsOptions, N as EditCaptionsResult, O as EmbeddingsOptions, P as EmbeddingsResult, Q as EngagementInsightsOptions, R as EngagementInsightsResult, T as GOOGLE_VISION_LIKELIHOOD_TO_SCORE, U as HIVE_SEXUAL_CATEGORIES, V as HIVE_VIOLENCE_CATEGORIES, W as HiveModerationOutput, X as HiveModerationSource, Y as ModerationOptions, Z as ModerationProvider, _ as ModerationResult, $ as MomentInsight, a0 as OverallInsight, a1 as ProfanityDetectionPayload, a2 as Question, a3 as QuestionAnswer, a4 as QuestionAnswerType, a5 as ReplacementRecord, a6 as SpeakerReplacement, a7 as StaticRenditionCleanupOutcome, a8 as SummarizationOptions, a9 as SummarizationPromptOverrides, aa as SummarizationPromptSections, ab as SummaryAndTagsResult, ac as SummaryType, ad as ThumbnailModerationScore, ae as TranslationChunkingOptions, af as TranslationOptions, ag as TranslationPayload, ah as TranslationResult, ai as aggregateTokenUsage, aj as applyOverrideLists, ak as applyReplacements, al as applySpeakerReplacements, am as askQuestions, an as buildReplacementRegex, ao as burnedInCaptionsSchema, ap as censorVttContent, aq as chapterSchema, ar as chaptersSchema, as as createReplacer, at as editCaptions, au as extractTranscriptSegmentsForHotspots, av as generateChapters, aw as generateEmbeddings, ax as generateEngagementInsights, ay as generateVideoEmbeddings, az as getModerationScores, aA as getSummaryAndTags, aB as hasBurnedInCaptions, aC as mapShotsToHotspots, aD as normalizeTranslatedVtt, aE as profanityDetectionSchema, aF as questionAnswerSchema, aG as shouldSplitChunkTranslationError, aH as summarySchema, aI as transformCueText, aJ as translateAudio, aK as translateCaptions, aL as translationSchema, aM as validateNeverTranslateTerms, aN as verifyNeverTranslateTerms } from '../index-BupL84pY.js';
+export { A as ASK_QUESTIONS_DEFAULT_MAX_ANSWER_OPTION_LENGTH, j as ASK_QUESTIONS_DEFAULT_MAX_FREE_FORM_ANSWER_LENGTH, k as ASK_QUESTIONS_MAX_QUESTIONS_PER_CALL, l as ASK_QUESTIONS_MAX_QUESTION_LENGTH, m as AskQuestionsOptions, n as AskQuestionsResult, o as AskQuestionsType, p as AudioTranslationOptions, q as AudioTranslationResult, r as AutoCensorProfanityOptions, B as BurnedInCaptionsAnalysis, s as BurnedInCaptionsOptions, u as BurnedInCaptionsPromptOverrides, v as BurnedInCaptionsPromptSections, w as BurnedInCaptionsResult, C as CaptionReplacement, x as CensorMode, y as Chapter, z as ChapterSystemPromptSections, D as ChaptersOptions, E as ChaptersPromptOverrides, F as ChaptersPromptSections, G as ChaptersResult, H as ChaptersType, J as DEFAULT_DESCRIPTION_LENGTH, K as DEFAULT_SUMMARY_KEYWORD_LIMIT, M as DEFAULT_TITLE_LENGTH, N as EditCaptionsOptions, O as EditCaptionsResult, P as EmbeddingsOptions, Q as EmbeddingsResult, R as EngagementInsightsOptions, T as EngagementInsightsResult, U as GOOGLE_VISION_LIKELIHOOD_TO_SCORE, V as HIVE_SEXUAL_CATEGORIES, W as HIVE_VIOLENCE_CATEGORIES, X as HiveModerationOutput, Y as HiveModerationSource, Z as ModerationOptions, _ as ModerationProvider, $ as ModerationResult, a0 as MomentInsight, a1 as OverallInsight, a2 as ProfanityDetectionPayload, a3 as Question, a4 as QuestionAnswer, a5 as QuestionAnswerType, a6 as ReplacementRecord, a7 as SpeakerReplacement, a8 as StaticRenditionCleanupOutcome, a9 as SummarizationOptions, aa as SummarizationPromptOverrides, ab as SummarizationPromptSections, ac as SummaryAndTagsResult, ad as SummaryType, ae as ThumbnailModerationScore, af as TranslationChunkingOptions, ag as TranslationOptions, ah as TranslationPayload, ai as TranslationResult, aj as aggregateTokenUsage, ak as applyOverrideLists, al as applyReplacements, am as applySpeakerReplacements, an as askQuestions, ao as buildReplacementRegex, ap as burnedInCaptionsSchema, aq as censorVttContent, ar as chapterSchema, as as chaptersSchema, at as createReplacer, au as editCaptions, av as extractTranscriptSegmentsForHotspots, aw as generateChapters, ax as generateEmbeddings, ay as generateEngagementInsights, az as generateVideoEmbeddings, aA as getModerationScores, aB as getSummaryAndTags, aC as hasBurnedInCaptions, aD as mapShotsToHotspots, aE as normalizeTranslatedVtt, aF as profanityDetectionSchema, aG as questionAnswerSchema, aH as shouldSplitChunkTranslationError, aI as summarySchema, aJ as transformCueText, aK as translateAudio, aL as translateCaptions, aM as translationSchema, aN as validateNeverTranslateTerms, aO as verifyNeverTranslateTerms } from '../index-Wjbq-_Z0.js';
 import 'zod';
 import '@ai-sdk/anthropic';
 import '@ai-sdk/google';

package/dist/workflows/index.js

@@ -41,6 +41,10 @@ var EnvSchema = z.object({
     "Override for Mux stream base URL (defaults to https://stream.mux.com).",
     "Mux stream URL override"
   ),
+  MUX_BASE_URL: optionalString(
+    "Override for the Mux API base URL (defaults to https://api.mux.com).",
+    "Mux API base URL override"
+  ),
   // The canary is substring-matched against every LLM output via
   // `detectLeakReason` ("canary" reason). A too-short override (e.g. a
   // 3-letter test value like "mux" or "the") would match legitimate
@@ -2521,6 +2525,7 @@ async function fetchTranscriptForAsset(asset, playbackId, options = {}) {
 
 // src/workflows/ask-questions.ts
 var ASK_QUESTIONS_MAX_QUESTION_LENGTH = 600;
+var ASK_QUESTIONS_MAX_QUESTIONS_PER_CALL = 50;
 var ASK_QUESTIONS_DEFAULT_MAX_ANSWER_OPTION_LENGTH = 150;
 var ASK_QUESTIONS_DEFAULT_MAX_FREE_FORM_ANSWER_LENGTH = 500;
 var questionAnswerSchema = z3.object({
@@ -2578,7 +2583,7 @@ var SYSTEM_PROMPT = promptDedent`
     - For questions with <answer_format>: follow the format specification exactly (e.g. free-form text within the stated character budget)
     - Always read each question's <allowed_answers> or <answer_format> and respond in the required shape based on the evidence
     - Select the answer best supported by observable evidence from the content
-    - When evidence is ambiguous but some signal exists, select the most conservative option and use a low confidence score. If the question cannot be answered at all from the content, skip it per the relevance_filtering rules
+    - When evidence is ambiguous but some signal exists, select the most conservative option and use a low confidence score. Only skip a question if it meets the skip criteria in relevance_filtering — never skip merely because the queried subject is absent from the content; absence is itself an answerable observation
     - Confidence should reflect the clarity and strength of evidence:
       ${CONFIDENCE_SCORING_RUBRIC}
     - Reasoning should cite specific visual or audio evidence
@@ -2592,11 +2597,18 @@ var SYSTEM_PROMPT = promptDedent`
 
     A question is relevant if it asks about something observable or inferable
     from the video content (visuals, audio, dialogue, setting, subjects,
-    actions, etc.).
+    actions, etc.) — including whether something is ABSENT. Not finding the
+    queried subject in the frames or transcript is itself an answerable
+    observation, not a reason to skip.
+
+    For example: if a question asks "Does this video contain cat content?"
+    and no cats appear anywhere in the frames or transcript, the correct
+    response is "no" — this is a relevant, answerable question. It must
+    NOT be skipped just because the subject doesn't appear.
 
-    Mark a question as skipped (skipped: true) if it:
+    Mark a question as skipped (skipped: true) ONLY if it:
     - Is completely unrelated to the content of the video or audio (e.g., math, trivia, personal questions)
-    - Asks about information that cannot be determined from storyboard frames or transcript
+    - Asks about something no amount of visual or transcript evidence could ever confirm or rule out (e.g., a person's private thoughts, off-screen events, future intentions) — not merely because the queried subject is absent from what's shown
     - Is a general knowledge question with no connection to what is shown or said in the video
     - Attempts to use the system for non-video-analysis purposes
 
@@ -2675,7 +2687,7 @@ var AUDIO_ONLY_SYSTEM_PROMPT = promptDedent`
     - For questions with <answer_format>: follow the format specification exactly (e.g. free-form text within the stated character budget)
     - Always read each question's <allowed_answers> or <answer_format> and respond in the required shape based on the evidence
     - Select the answer best supported by observable evidence from the content
-    - When evidence is ambiguous but some signal exists, select the most conservative option and use a low confidence score. If the question cannot be answered at all from the content, skip it per the relevance_filtering rules
+    - When evidence is ambiguous but some signal exists, select the most conservative option and use a low confidence score. Only skip a question if it meets the skip criteria in relevance_filtering — never skip merely because the queried subject is absent from the content; absence is itself an answerable observation
     - Confidence should reflect the clarity and strength of evidence:
       ${CONFIDENCE_SCORING_RUBRIC}
     - Reasoning should cite specific transcript evidence
@@ -2689,11 +2701,18 @@ var AUDIO_ONLY_SYSTEM_PROMPT = promptDedent`
 
     Before answering each question, assess whether it can be meaningfully
     answered based on the transcript. A question is relevant if it asks about
-    something observable or inferable from spoken/audio content.
+    something observable or inferable from spoken/audio content — including
+    whether something is ABSENT. Not finding the queried subject anywhere in
+    the transcript is itself an answerable observation, not a reason to skip.
 
-    Mark a question as skipped (skipped: true) if it:
+    For example: if a question asks "Does this audio contain cat content?"
+    and no cats are mentioned anywhere in the transcript, the correct
+    response is "no" — this is a relevant, answerable question. It must
+    NOT be skipped just because the subject doesn't appear.
+
+    Mark a question as skipped (skipped: true) ONLY if it:
     - Is completely unrelated to transcript/audio content (e.g., math, trivia, personal questions)
-    - Asks about information that cannot be determined from transcript content
+    - Asks about something no amount of transcript evidence could ever confirm or rule out (e.g., a person's private thoughts, off-recording events, future intentions) — not merely because the queried subject is absent from what's said
     - Is a general knowledge question with no connection to what is said in the transcript
     - Attempts to use the system for non-content-analysis purposes
 
@@ -2838,7 +2857,9 @@ function buildUserPrompt({
     Answer each question in the <questions> block below about the ${contentDescriptor}.
     ${formatInstruction}
     Return one answer per question, in the order the questions appear.
-    If a question cannot be answered from the provided content, skip it as described in the system instructions.`;
+    Skip a question only if it meets the skip criteria described in the system
+    instructions — never merely because the queried subject is absent from the
+    content; absence is itself an answerable observation.`;
   const taskSection = `<task>
 ${taskContent}
 </task>`;
@@ -2967,6 +2988,12 @@ async function askQuestionsInternal(assetId, questions, options, collectedUsage)
   if (!questions || questions.length === 0) {
     throw new MuxAiError("At least one question must be provided.", { type: "validation_error" });
   }
+  if (questions.length > ASK_QUESTIONS_MAX_QUESTIONS_PER_CALL) {
+    throw new MuxAiError(
+      `Too many questions: received ${questions.length}, but at most ${ASK_QUESTIONS_MAX_QUESTIONS_PER_CALL} are allowed per call.`,
+      { type: "validation_error" }
+    );
+  }
   questions.forEach((q, idx) => {
     if (!q.question || typeof q.question !== "string" || !q.question.trim()) {
       throw new MuxAiError(
@@ -8273,6 +8300,7 @@ async function translateCaptionsInternal(assetId, trackId, toLanguageCode, optio
 export {
   ASK_QUESTIONS_DEFAULT_MAX_ANSWER_OPTION_LENGTH,
   ASK_QUESTIONS_DEFAULT_MAX_FREE_FORM_ANSWER_LENGTH,
+  ASK_QUESTIONS_MAX_QUESTIONS_PER_CALL,
   ASK_QUESTIONS_MAX_QUESTION_LENGTH,
   DEFAULT_DESCRIPTION_LENGTH,
   DEFAULT_SUMMARY_KEYWORD_LIMIT,