remove from omni api

2026-10-20 17:02:53 -04:00 · 2026-08-06 17:38:01 -04:00
parent c21463aa1a
commit 55a7aa35b5
7 changed files with 100 additions and 425 deletions
@@ -1,150 +0,0 @@
-//! Enqueue flow for first-party Minimax H3 (Turbo / Ultra) video generation.
-//!
-//! These models run on our own GPU inference, so unlike the provider-backed
-//! models there is no external generation call: we bill (Ultra only), write
-//! the prompt + job records, and return. A scheduler will pick the pending
-//! jobs up later.
-
-use std::convert::TryFrom;
-
-use actix_web::web::Json;
-use log::{error, info};
-use sqlx::Acquire;
-
-use artcraft_api_defs::omni_api::generate_requests::omni_api_video_generate_request::OmniApiVideoGenerateRequest;
-use artcraft_api_defs::omni_gen::generate_response::omni_gen_video_generate_response::OmniGenVideoGenerateResponse;
-use enums::common::generation::common_model_type::CommonModelType;
-use enums::common::generation::common_video_model::CommonVideoModel;
-use enums::common::platform_type::PlatformType;
-use enums::common::visibility::Visibility;
-use mysql_queries::queries::generic_inference::first_party::minimax_h3::insert_generic_inference_job_for_first_party_minimax_h3_with_apriori_job_token::{
-  insert_generic_inference_job_for_first_party_minimax_h3_with_apriori_job_token,
-  FirstPartyMinimaxH3Model,
-  InsertGenericInferenceForFirstPartyMinimaxH3WithAprioriJobTokenArgs,
-};
-use tokens::tokens::non_unique::debug_logs_event_token::DebugLogEventToken;
-use tokens::tokens::users::UserToken;
-
-use crate::http_server::common_responses::common_web_error::CommonWebError;
-use crate::http_server::endpoints::omni_api::generate::video::first_party_minimax_h3::minimax_h3_ultra_cost::estimate_minimax_h3_ultra_cost_usd_cents;
-use crate::http_server::endpoints::omni_api::generate::video::helpers::bill_wallet::bill_wallet;
-use crate::http_server::endpoints::omni_api::generate::video::helpers::write_prompt_records::{
-  write_prompt_records, WritePromptRecordsArgs,
-};
-
-pub struct EnqueueFirstPartyMinimaxH3JobArgs<'a> {
-  pub minimax_model: FirstPartyMinimaxH3Model,
-  pub request: &'a OmniApiVideoGenerateRequest,
-  pub user_token: &'a UserToken,
-  pub maybe_prompt_model_type: Option<CommonModelType>,
-  pub idempotency_token: &'a str,
-  pub ip_address: &'a str,
-  pub debug_log_event_token: &'a DebugLogEventToken,
-  pub mysql_connection: sqlx::pool::PoolConnection<sqlx::MySql>,
-}
-
-/// Map the requested video model to a first-party Minimax H3 tier.
-///
-/// Returns `None` for every other model — including the bare `MinimaxH3`,
-/// which stays on its existing third-party (Fal) route.
-pub fn first_party_minimax_h3_model(
-  maybe_model: Option<CommonVideoModel>,
-) -> Option<FirstPartyMinimaxH3Model> {
-  match maybe_model {
-    Some(CommonVideoModel::MinimaxH3Turbo) => Some(FirstPartyMinimaxH3Model::Turbo),
-    Some(CommonVideoModel::MinimaxH3Ultra) => Some(FirstPartyMinimaxH3Model::Ultra),
-    _ => None,
-  }
-}
-
-pub async fn enqueue_first_party_minimax_h3_job(
-  args: EnqueueFirstPartyMinimaxH3JobArgs<'_>,
-) -> Result<Json<OmniGenVideoGenerateResponse>, CommonWebError> {
-  let EnqueueFirstPartyMinimaxH3JobArgs {
-    minimax_model,
-    request,
-    user_token,
-    maybe_prompt_model_type,
-    idempotency_token,
-    ip_address,
-    debug_log_event_token,
-    mut mysql_connection,
-  } = args;
-
-  // System credits = cents (same convention as the v2 pipeline).
-  let maybe_system_cost_usd_cents: Option<u64> = match minimax_model {
-    FirstPartyMinimaxH3Model::Turbo => {
-      // Turbo is free: no wallet charge.
-      // TODO: Gate free Turbo generations on how many jobs the user has
-      //  already submitted. The count queries and checks land with the
-      //  scheduler work.
-      None
-    }
-    FirstPartyMinimaxH3Model::Ultra => {
-      Some(estimate_minimax_h3_ultra_cost_usd_cents(request))
-    }
-  };
-
-  info!("First-party Minimax H3 ({:?}) estimated cost: {:?} credits",
-    minimax_model, maybe_system_cost_usd_cents);
-
-  // Bill the wallet (no-op deduction for the free Turbo tier) and generate
-  // the apriori job token.
-  let billing = bill_wallet(
-    user_token,
-    maybe_system_cost_usd_cents.unwrap_or(0),
-    &mut mysql_connection,
-  ).await?;
-
-  let mut transaction = mysql_connection.begin().await.map_err(|err| {
-    error!("Error starting MySQL transaction: {:?}", err);
-    CommonWebError::from_error(err)
-  })?;
-
-  let prompt_token = write_prompt_records(WritePromptRecordsArgs {
-    request,
-    user_token,
-    maybe_prompt_model_type,
-    ip_address,
-    transaction: &mut transaction,
-  }).await;
-
-  let maybe_system_cost_u32 = maybe_system_cost_usd_cents
-    .and_then(|cents| u32::try_from(cents).ok());
-
-  let job_token = insert_generic_inference_job_for_first_party_minimax_h3_with_apriori_job_token(
-    InsertGenericInferenceForFirstPartyMinimaxH3WithAprioriJobTokenArgs {
-      minimax_model,
-      uuid_idempotency_token: idempotency_token,
-      apriori_job_token: &billing.apriori_job_token,
-      maybe_inference_args: None,
-      maybe_prompt_token: prompt_token.as_ref(),
-      maybe_wallet_ledger_entry_token: billing.maybe_wallet_ledger_entry_token.as_ref(),
-      maybe_system_cost_credits: maybe_system_cost_u32,
-      maybe_system_cost_usd_cents: maybe_system_cost_u32,
-      maybe_creator_user_token: Some(user_token),
-      maybe_avt_token: None, // AVT tokens are web-session only; API-key callers have none.
-      creator_ip_address: ip_address,
-      creator_set_visibility: Visibility::Public,
-      // Omni API requests are always API-key authenticated.
-      maybe_platform_type: Some(PlatformType::ApiKey),
-      maybe_debug_log_event_token: Some(debug_log_event_token),
-      mysql_executor: &mut *transaction,
-      phantom: Default::default(),
-    }
-  ).await.map_err(|err| {
-    error!("Error inserting first-party Minimax H3 job: {:?}", err);
-    CommonWebError::from_error(err)
-  })?;
-
-  transaction.commit().await.map_err(|err| {
-    error!("Error committing transaction: {:?}", err);
-    CommonWebError::from_error(err)
-  })?;
-
-  Ok(Json(OmniGenVideoGenerateResponse {
-    success: true,
-    inference_job_token: job_token.clone(),
-    all_job_tokens: vec![job_token],
-  }))
-}
@@ -1,101 +0,0 @@
-//! Pricing for first-party Minimax H3 Ultra.
-//!
-//! TODO(2026-08-06): Placeholder rates copied from the existing Minimax H3
-//! pricing. Finalize once first-party GPU inference costs are known.
-
-use artcraft_api_defs::omni_api::generate_requests::omni_api_video_generate_request::OmniApiVideoGenerateRequest;
-use enums::common::generation::common_resolution::CommonResolution;
-
-/// Per-second rates in hundredths of a US cent.
-const LOW_RES_RATE_HUNDREDTH_CENTS_PER_SEC: u64 = 1_840;
-const HIGH_RES_RATE_HUNDREDTH_CENTS_PER_SEC: u64 = 2_990;
-
-/// Minimax H3 defaults None → 5s.
-const DEFAULT_DURATION_SECONDS: u64 = 5;
-
-/// Estimate the user-facing Minimax H3 Ultra cost in USD cents.
-/// System credits = cents (same convention as the v2 pipeline).
-pub fn estimate_minimax_h3_ultra_cost_usd_cents(request: &OmniApiVideoGenerateRequest) -> u64 {
-  let duration_seconds = request.duration_seconds
-    .map(u64::from)
-    .unwrap_or(DEFAULT_DURATION_SECONDS);
-
-  let rate = if is_2k(request.resolution) {
-    HIGH_RES_RATE_HUNDREDTH_CENTS_PER_SEC
-  } else {
-    LOW_RES_RATE_HUNDREDTH_CENTS_PER_SEC
-  };
-
-  // Round up to the next whole cent.
-  (rate * duration_seconds).div_ceil(100)
-}
-
-/// Minimax H3 renders 768P or 2K; 1080p and above land on the 2K tier, while
-/// 720p and below render at 768P. The default (unset) resolution is 2K.
-fn is_2k(resolution: Option<CommonResolution>) -> bool {
-  match resolution {
-    None => true, // defaults to 2K
-    Some(CommonResolution::TenEightyP)
-    | Some(CommonResolution::OneK)
-    | Some(CommonResolution::TwoK)
-    | Some(CommonResolution::ThreeK)
-    | Some(CommonResolution::FourK) => true,
-    Some(CommonResolution::HalfK)
-    | Some(CommonResolution::FourEightyP)
-    | Some(CommonResolution::SevenTwentyP) => false,
-  }
-}
-
-#[cfg(test)]
-mod tests {
-  use super::*;
-
-  #[test]
-  fn default_resolution_5s_is_150() {
-    assert_eq!(cost_cents(None, None), 150);
-  }
-
-  #[test]
-  fn low_res_5s_is_92() {
-    assert_eq!(cost_cents(Some(5), Some(CommonResolution::SevenTwentyP)), 92);
-  }
-
-  #[test]
-  fn high_res_10s_is_299() {
-    assert_eq!(cost_cents(Some(10), Some(CommonResolution::TenEightyP)), 299);
-  }
-
-  #[test]
-  fn odd_duration_rounds_up_to_whole_cents() {
-    // 7s high res: 2990 × 7 = 20930 hundredth-cents → 210 cents.
-    assert_eq!(cost_cents(Some(7), None), 210);
-  }
-
-  fn cost_cents(duration_seconds: Option<u16>, resolution: Option<CommonResolution>) -> u64 {
-    let request = OmniApiVideoGenerateRequest {
-      idempotency_token: None,
-      model: None,
-      prompt: None,
-      negative_prompt: None,
-      start_frame_image_media_token: None,
-      start_frame_image_url: None,
-      end_frame_image_media_token: None,
-      end_frame_image_url: None,
-      reference_image_media_tokens: None,
-      reference_image_urls: None,
-      reference_video_media_tokens: None,
-      reference_video_urls: None,
-      reference_audio_media_tokens: None,
-      reference_audio_urls: None,
-      reference_character_tokens: None,
-      resolution,
-      aspect_ratio: None,
-      bitrate: None,
-      quality: None,
-      duration_seconds,
-      video_batch_count: None,
-      generate_audio: None,
-    };
-    estimate_minimax_h3_ultra_cost_usd_cents(&request)
-  }
-}
@@ -1,8 +0,0 @@
-//! First-party Minimax H3 (Turbo / Ultra) video generation.
-//!
-//! These models are fulfilled by our own GPU inference rather than an
-//! external provider, so enqueueing only writes database records — a
-//! scheduler runs the jobs later.
-
-pub mod enqueue_first_party_minimax_h3_job;
-pub mod minimax_h3_ultra_cost;
@@ -4,4 +4,3 @@ pub(crate) mod hydrate_router_request;
 pub mod pipeline_result;
 pub mod resolve_kinovi_character_ids;
 pub mod resolve_media_tokens_to_urls;
-pub mod write_prompt_records;
@@ -1,132 +0,0 @@
-//! Shared prompt bookkeeping for the omni-api video generate flows.
-//!
-//! Writes the prompt row and its context items (start/end frames, image and
-//! video references) inside the caller's transaction. Best-effort: failures
-//! are logged and yield `None` — job enqueue must not fail because prompt
-//! bookkeeping did.
-
-use log::warn;
-use sqlx::{MySql, Transaction};
-
-use artcraft_api_defs::omni_api::generate_requests::omni_api_video_generate_request::OmniApiVideoGenerateRequest;
-use enums::by_table::prompt_context_items::prompt_context_semantic_type::PromptContextSemanticType;
-use enums::by_table::prompts::prompt_type::PromptType;
-use enums::common::generation::common_generation_mode::CommonGenerationMode;
-use enums::common::generation::common_model_type::CommonModelType;
-use enums::common::generation_provider::GenerationProvider;
-use mysql_queries::queries::prompt_context_items::insert_batch_prompt_context_items::{
-  insert_batch_prompt_context_items, InsertBatchArgs, PromptContextItem,
-};
-use mysql_queries::queries::prompts::insert_prompt::{insert_prompt, InsertPromptArgs};
-use tokens::tokens::prompts::PromptToken;
-use tokens::tokens::users::UserToken;
-
-pub struct WritePromptRecordsArgs<'a, 'tx> {
-  pub request: &'a OmniApiVideoGenerateRequest,
-  pub user_token: &'a UserToken,
-  pub maybe_prompt_model_type: Option<CommonModelType>,
-  pub ip_address: &'a str,
-  pub transaction: &'a mut Transaction<'tx, MySql>,
-}
-
-pub async fn write_prompt_records(args: WritePromptRecordsArgs<'_, '_>) -> Option<PromptToken> {
-  let WritePromptRecordsArgs {
-    request,
-    user_token,
-    maybe_prompt_model_type,
-    ip_address,
-    transaction,
-  } = args;
-
-  let prompt_token = match insert_prompt(InsertPromptArgs {
-    maybe_apriori_prompt_token: None,
-    prompt_type: PromptType::ArtcraftApp,
-    maybe_creator_user_token: Some(user_token),
-    maybe_model_type: maybe_prompt_model_type,
-    maybe_generation_provider: Some(GenerationProvider::Artcraft),
-    maybe_positive_prompt: request.prompt.as_deref(),
-    maybe_negative_prompt: request.negative_prompt.as_deref(),
-    maybe_other_args: None,
-    maybe_generation_mode: Some(determine_generation_mode(request)),
-    maybe_aspect_ratio: request.aspect_ratio,
-    maybe_resolution: request.resolution,
-    maybe_bitrate: request.bitrate,
-    maybe_batch_count: request.video_batch_count.map(|c| c as u8),
-    maybe_generate_audio: request.generate_audio,
-    maybe_duration_seconds: request.duration_seconds.map(|d| d as u32),
-    creator_ip_address: ip_address,
-    mysql_executor: &mut **transaction,
-    phantom: Default::default(),
-  }).await {
-    Ok(token) => Some(token),
-    Err(err) => {
-      warn!("Error inserting prompt: {:?}", err);
-      None
-    }
-  };
-
-  if let Some(token) = prompt_token.as_ref() {
-    let mut context_items = Vec::new();
-
-    if let Some(media_token) = &request.start_frame_image_media_token {
-      context_items.push(PromptContextItem {
-        media_token: media_token.clone(),
-        context_semantic_type: PromptContextSemanticType::VidStartFrame,
-      });
-    }
-    if let Some(media_token) = &request.end_frame_image_media_token {
-      context_items.push(PromptContextItem {
-        media_token: media_token.clone(),
-        context_semantic_type: PromptContextSemanticType::VidEndFrame,
-      });
-    }
-    if let Some(ref_tokens) = &request.reference_image_media_tokens {
-      for media_token in ref_tokens {
-        context_items.push(PromptContextItem {
-          media_token: media_token.clone(),
-          context_semantic_type: PromptContextSemanticType::Imgref,
-        });
-      }
-    }
-    if let Some(ref_tokens) = &request.reference_video_media_tokens {
-      for media_token in ref_tokens {
-        context_items.push(PromptContextItem {
-          media_token: media_token.clone(),
-          context_semantic_type: PromptContextSemanticType::VidRef,
-        });
-      }
-    }
-
-    if !context_items.is_empty() {
-      if let Err(err) = insert_batch_prompt_context_items(InsertBatchArgs {
-        prompt_token: token.clone(),
-        items: context_items,
-        transaction,
-      }).await {
-        warn!("Error inserting batch prompt context items: {:?}", err);
-      }
-    }
-  }
-
-  prompt_token
-}
-
-fn determine_generation_mode(request: &OmniApiVideoGenerateRequest) -> CommonGenerationMode {
-  let has_keyframe = request.start_frame_image_media_token.is_some()
-    || request.end_frame_image_media_token.is_some();
-
-  if has_keyframe {
-    return CommonGenerationMode::Keyframe;
-  }
-
-  let has_reference = request.reference_image_media_tokens.as_ref().is_some_and(|t| !t.is_empty())
-    || request.reference_video_media_tokens.as_ref().is_some_and(|t| !t.is_empty())
-    || request.reference_audio_media_tokens.as_ref().is_some_and(|t| !t.is_empty())
-    || request.reference_character_tokens.as_ref().is_some_and(|t| !t.is_empty());
-
-  if has_reference {
-    return CommonGenerationMode::Reference;
-  }
-
-  CommonGenerationMode::Text
-}
@@ -1,6 +1,5 @@
 pub mod omni_api_video_generate_handler;
 pub mod check_request;
-pub mod first_party_minimax_h3;
 pub mod helpers;
 pub mod ingest_url_inputs;
 pub mod pipeline_v2;
@@ -10,14 +10,22 @@ use artcraft_api_defs::omni_api::generate_requests::omni_api_video_generate_requ
 use artcraft_api_defs::omni_gen::generate_response::omni_gen_video_generate_response::OmniGenVideoGenerateResponse;
 use artcraft_router::generate::generate_video::generate_video_response::GenerateVideoResponse;
 use enums::by_table::debug_logs::debug_log_type::DebugLogType;
+use enums::by_table::prompt_context_items::prompt_context_semantic_type::PromptContextSemanticType;
+use enums::by_table::prompts::prompt_type::PromptType;
+use enums::common::generation::common_generation_mode::CommonGenerationMode;
 use enums::common::generation::common_model_type::CommonModelType;
 use enums::common::generation::common_video_model::CommonVideoModel;
+use enums::common::generation_provider::GenerationProvider;
 use enums::common::platform_type::PlatformType;
 use http_server_common::request::get_request_ip::get_request_ip;
 use enums::by_table::debug_logs::debug_log_level::DebugLogLevel;
 use mysql_queries::queries::debug_logs::insert_debug_log::{insert_debug_log, InsertDebugLogArgs};
 use mysql_queries::queries::generic_inference::api_providers::kinovi_web::insert_generic_inference_job_for_kinovi_web_queue_with_apriori_job_token::KinoviVersion;
 use mysql_queries::queries::idepotency_tokens::insert_idempotency_token::insert_idempotency_token;
+use mysql_queries::queries::prompt_context_items::insert_batch_prompt_context_items::{
+  insert_batch_prompt_context_items, InsertBatchArgs, PromptContextItem,
+};
+use mysql_queries::queries::prompts::insert_prompt::{insert_prompt, InsertPromptArgs};
 use tokens::tokens::characters::CharacterToken;
 use tokens::tokens::generic_inference_jobs::InferenceJobToken;
 use tokens::tokens::media_files::MediaFileToken;
@@ -27,9 +35,6 @@ use crate::http_server::common_responses::common_web_error::CommonWebError;
 use crate::http_server::endpoints::generate::common::generation_debug_logs::GenerationDebugLogContext;
 use crate::http_server::endpoints::generate::common::payments_error_test::payments_error_test;
 use crate::http_server::endpoints::omni_api::generate::video::check_request::check_request;
-use crate::http_server::endpoints::omni_api::generate::video::first_party_minimax_h3::enqueue_first_party_minimax_h3_job::{
-  enqueue_first_party_minimax_h3_job, first_party_minimax_h3_model, EnqueueFirstPartyMinimaxH3JobArgs,
-};
 use crate::http_server::endpoints::omni_api::generate::video::ingest_url_inputs::ingest_url_inputs;
 use crate::http_server::endpoints::omni_api::generate::video::helpers::hydrate_router_request::hydrate_to_router_request;
 use crate::http_server::endpoints::omni_api::generate::video::helpers::resolve_kinovi_character_ids::resolve_kinovi_character_ids;
@@ -38,9 +43,6 @@ use crate::http_server::endpoints::omni_api::generate::video::insert_db_job::ins
 use crate::http_server::endpoints::omni_api::generate::video::insert_db_job::insert_grok_api_job::{insert_grok_api_job, InsertGrokApiJobArgs};
 use crate::http_server::endpoints::omni_api::generate::video::insert_db_job::insert_kinovi_web_jobs::{insert_kinovi_web_jobs, InsertKinoviWebJobsArgs};
 use crate::http_server::endpoints::omni_api::generate::video::insert_db_job::shared_job_args::SharedJobArgs;
-use crate::http_server::endpoints::omni_api::generate::video::helpers::write_prompt_records::{
-  write_prompt_records, WritePromptRecordsArgs,
-};
 use crate::http_server::endpoints::omni_api::shared_utils::kinovi_account::KinoviAccount;
 use crate::http_server::endpoints::omni_api::generate::video::pipeline_v2::run_pipeline_v2::{run_pipeline_v2, RunPipelineV2Args};
 use crate::http_server::endpoints::omni_api::shared_utils::video::validate_video_request::validate_video_request;
@@ -216,24 +218,6 @@ pub async fn omni_api_video_generate_handler(
    warn!("Failed to insert HTTP request debug log: {:?}", err);
  }

-  // ==================== FIRST-PARTY MINIMAX H3 (TURBO / ULTRA) ==================== //
-
-  // These models run on our own GPU inference — no provider call is made.
-  // The module bills the wallet (Ultra only), writes the prompt + job
-  // records, and returns; a scheduler picks the pending jobs up later.
-  if let Some(minimax_model) = first_party_minimax_h3_model(request.model) {
-    return enqueue_first_party_minimax_h3_job(EnqueueFirstPartyMinimaxH3JobArgs {
-      minimax_model,
-      request: &request,
-      user_token,
-      maybe_prompt_model_type,
-      idempotency_token: &idempotency_token,
-      ip_address: &ip_address,
-      debug_log_event_token: &debug_log_event_token,
-      mysql_connection,
-    }).await;
-  }
-
  // NB: The pipeline takes over the connection for its remaining pre-request DB writes (billing,
  // outbound provider request debug log) and releases it before the (slow, external) generation
  // call — holding a pool slot across that call is what starves the pool and causes PoolTimedOut
@@ -297,15 +281,79 @@ pub async fn omni_api_video_generate_handler(
    CommonWebError::from_error(err)
  })?;

-  // -- Prompt + context items --
+  // -- Prompt --

-  let prompt_token = write_prompt_records(WritePromptRecordsArgs {
-    request: &request,
-    user_token,
-    maybe_prompt_model_type,
-    ip_address: &ip_address,
-    transaction: &mut transaction,
-  }).await;
+  let prompt_token = match insert_prompt(InsertPromptArgs {
+    maybe_apriori_prompt_token: None,
+    prompt_type: PromptType::ArtcraftApp,
+    maybe_creator_user_token: Some(user_token),
+    maybe_model_type: maybe_prompt_model_type,
+    maybe_generation_provider: Some(GenerationProvider::Artcraft),
+    maybe_positive_prompt: request.prompt.as_deref(),
+    maybe_negative_prompt: request.negative_prompt.as_deref(),
+    maybe_other_args: None,
+    maybe_generation_mode: Some(determine_generation_mode(&request)),
+    maybe_aspect_ratio: request.aspect_ratio,
+    maybe_resolution: request.resolution,
+    maybe_bitrate: request.bitrate,
+    maybe_batch_count: request.video_batch_count.map(|c| c as u8),
+    maybe_generate_audio: request.generate_audio,
+    maybe_duration_seconds: request.duration_seconds.map(|d| d as u32),
+    creator_ip_address: &ip_address,
+    mysql_executor: &mut *transaction,
+    phantom: Default::default(),
+  }).await {
+    Ok(token) => Some(token),
+    Err(err) => {
+      warn!("Error inserting prompt: {:?}", err);
+      None
+    }
+  };
+
+  // -- Prompt context items --
+
+  if let Some(token) = prompt_token.as_ref() {
+    let mut context_items = Vec::new();
+
+    if let Some(media_token) = &request.start_frame_image_media_token {
+      context_items.push(PromptContextItem {
+        media_token: media_token.clone(),
+        context_semantic_type: PromptContextSemanticType::VidStartFrame,
+      });
+    }
+    if let Some(media_token) = &request.end_frame_image_media_token {
+      context_items.push(PromptContextItem {
+        media_token: media_token.clone(),
+        context_semantic_type: PromptContextSemanticType::VidEndFrame,
+      });
+    }
+    if let Some(ref_tokens) = &request.reference_image_media_tokens {
+      for media_token in ref_tokens {
+        context_items.push(PromptContextItem {
+          media_token: media_token.clone(),
+          context_semantic_type: PromptContextSemanticType::Imgref,
+        });
+      }
+    }
+    if let Some(ref_tokens) = &request.reference_video_media_tokens {
+      for media_token in ref_tokens {
+        context_items.push(PromptContextItem {
+          media_token: media_token.clone(),
+          context_semantic_type: PromptContextSemanticType::VidRef,
+        });
+      }
+    }
+
+    if !context_items.is_empty() {
+      if let Err(err) = insert_batch_prompt_context_items(InsertBatchArgs {
+        prompt_token: token.clone(),
+        items: context_items,
+        transaction: &mut transaction,
+      }).await {
+        warn!("Error inserting batch prompt context items: {:?}", err);
+      }
+    }
+  }

  // -- Inference job --

@@ -439,3 +487,23 @@ pub async fn omni_api_video_generate_handler(
    all_job_tokens,
  }))
 }
+
+fn determine_generation_mode(request: &OmniApiVideoGenerateRequest) -> CommonGenerationMode {
+  let has_keyframe = request.start_frame_image_media_token.is_some()
+    || request.end_frame_image_media_token.is_some();
+
+  if has_keyframe {
+    return CommonGenerationMode::Keyframe;
+  }
+
+  let has_reference = request.reference_image_media_tokens.as_ref().is_some_and(|t| !t.is_empty())
+    || request.reference_video_media_tokens.as_ref().is_some_and(|t| !t.is_empty())
+    || request.reference_audio_media_tokens.as_ref().is_some_and(|t| !t.is_empty())
+    || request.reference_character_tokens.as_ref().is_some_and(|t| !t.is_empty());
+
+  if has_reference {
+    return CommonGenerationMode::Reference;
+  }
+
+  CommonGenerationMode::Text
+}