Skip to content

Open-source LLM knowledge platform: turn raw documents into a queryable RAG, an autonomous reasoning agent, and a self-maintaining Wiki.

46.4 KB
Raw Blame
package service

import (
	"context"
	"crypto/sha256"
	"encoding/json"
	"errors"
	"fmt"
	"io"
	"os"
	"strings"
	"time"

	"github.com/Tencent/WeKnora/internal/application/repository"
	"github.com/Tencent/WeKnora/internal/application/service/retriever"
	"github.com/Tencent/WeKnora/internal/logger"
	"github.com/Tencent/WeKnora/internal/models/api"
	"github.com/Tencent/WeKnora/internal/models/embedding"
	"github.com/Tencent/WeKnora/internal/models/imageprep"
	"github.com/Tencent/WeKnora/internal/models/utils/ollama"
	"github.com/Tencent/WeKnora/internal/models/vlm"
	"github.com/Tencent/WeKnora/internal/tracing/langfuse"
	"github.com/Tencent/WeKnora/internal/types"
	"github.com/Tencent/WeKnora/internal/types/interfaces"
	secutils "github.com/Tencent/WeKnora/internal/utils"
	"github.com/google/uuid"
	"github.com/hibiken/asynq"
	"github.com/redis/go-redis/v9"
)

const (
	vlmOCRPrompt = "You are an OCR assistant. Transcribe only text that is visibly present in the image, " +
		"preserving its original language.\n" +
		"The image may be a screenshot, photograph, diagram, or document. For scattered labels, " +
		"buttons, scores, and other UI text, output plain text in reading order, one item per line. " +
		"Do not force this content into a table.\n" +
		"Use Markdown tables only when the image contains a real table with readable cells. Never " +
		"output empty table cells, invented layout, repeated filler, or descriptions of the image.\n" +
		"Preserve readable numbers and symbols. Do not guess unreadable text. Stop after all visible " +
		"text has been transcribed; repeat text only when it visibly appears more than once.\n" +
		"Output only the extracted text, without explanations, HTML tags, or code fences. If no text " +
		"is recognizable, reply only: No text content."
	vlmOCRScannedPDFPrompt = "<system_prompt>\n" +
		"You are an OCR and document layout extraction assistant. The input image is a page from a scanned PDF document.\n" +
		"Your task is to carefully extract all text and layout structure from the image, and output the result in pure Markdown format.\n" +
		"</system_prompt>\n\n" +
		"<instructions>\n" +
		"1. Ignore headers, footers, and page numbers.\n" +
		"2. Preserve the original document's paragraph and hierarchical structure as much as possible.\n" +
		"3. If there are tables, use Markdown table syntax to represent them.\n" +
		"4. If there are mathematical formulas, use LaTeX format wrapped in $ or $$.\n" +
		"5. Output ONLY the extracted text content. Do NOT include any HTML tags, reasoning, or unrelated comments.\n" +
		"6. If there is absolutely no recognizable text content in the image, reply ONLY with: No text content.\n" +
		"</instructions>"
)

func buildVLMCaptionPrompt(ctx context.Context, cfg types.VLMConfig) string {
	language := strings.TrimSpace(cfg.DescriptionLanguage)
	if language == "" {
		language = types.LanguageNameFromContext(ctx)
	}
	prompt := fmt.Sprintf("Provide a brief and concise description of the main content of the image in %s.", language)
	return types.AppendCustomPromptInstructions(prompt, cfg.CustomInstructions, "image_description")
}

// buildVLMOCRPrompt returns the system-owned OCR prompt for the given image
// source type. cfg is part of the signature and deliberately unused: knowledge
// base custom instructions must never reach the OCR prompt — free-form
// business rules compete with the "No text content" output contract and turn
// decorative images into poisoned, vectorized image_ocr chunks. Custom OCR
// guidance, if ever needed, deserves a dedicated OCR instruction field rather
// than reusing CustomInstructions.
func buildVLMOCRPrompt(sourceType string, _ types.VLMConfig) string {
	if sourceType == "scanned_pdf" {
		return vlmOCRScannedPDFPrompt
	}
	return vlmOCRPrompt
}

// ImageMultimodalService handles image:multimodal asynq tasks.
// It reads images from storage (via FileService for provider:// URLs),
// performs OCR and VLM caption, and creates child chunks.
type ImageMultimodalService struct {
	chunkService   interfaces.ChunkService
	modelService   interfaces.ModelService
	kbService      interfaces.KnowledgeBaseService
	knowledgeRepo  interfaces.KnowledgeRepository
	tenantRepo     interfaces.TenantRepository
	retrieveEngine interfaces.RetrieveEngineRegistry
	ownership      retriever.TenantStoreOwnership
	ollamaService  *ollama.OllamaService
	taskEnqueuer   interfaces.TaskEnqueuer
	redisClient    *redis.Client
	// fileSvc is the globally configured default FileService used as a fallback
	// when the tenant-scoped storage config cannot produce a usable service
	// (e.g. images were saved using the global MINIO_* env vars while the
	// tenant's StorageEngineConfig.MinIO is empty). Mirrors the write-side
	// fallback in knowledgeService.resolveFileService.
	fileSvc         interfaces.FileService
	storageResolver interfaces.StorageBackendResolver
	// resourceCatalog resolves resource:// references to their owning storage
	// backend so multimodal reads target the resource's real backend instead of
	// the knowledge base's currently configured one.
	resourceCatalog interfaces.ResourceCatalog

	// spanTracker records this image's subspan under the parent attempt's
	// multimodal stage. nil-safe — falls back to no-op via tracker().
	spanTracker SpanTracker

	// docReader is the document service the pipeline asks to straighten a
	// photographed page before OCR. Nil when this build has none (tests,
	// minimal wiring): the preprocess action then reports "unavailable" and
	// OCR simply reads the original bytes.
	docReader interfaces.DocumentReader
}

func NewImageMultimodalService(
	chunkService interfaces.ChunkService,
	modelService interfaces.ModelService,
	kbService interfaces.KnowledgeBaseService,
	knowledgeRepo interfaces.KnowledgeRepository,
	tenantRepo interfaces.TenantRepository,
	retrieveEngine interfaces.RetrieveEngineRegistry,
	ownership retriever.TenantStoreOwnership,
	ollamaService *ollama.OllamaService,
	taskEnqueuer interfaces.TaskEnqueuer,
	redisClient *redis.Client,
	fileSvc interfaces.FileService,
	storageResolver interfaces.StorageBackendResolver,
	resourceCatalog interfaces.ResourceCatalog,
	spanTracker SpanTracker,
	docReader interfaces.DocumentReader,
) interfaces.TaskHandler {
	return &ImageMultimodalService{
		chunkService:    chunkService,
		modelService:    modelService,
		kbService:       kbService,
		knowledgeRepo:   knowledgeRepo,
		tenantRepo:      tenantRepo,
		retrieveEngine:  retrieveEngine,
		ownership:       ownership,
		ollamaService:   ollamaService,
		taskEnqueuer:    taskEnqueuer,
		redisClient:     redisClient,
		fileSvc:         fileSvc,
		storageResolver: storageResolver,
		resourceCatalog: resourceCatalog,
		spanTracker:     spanTracker,
		docReader:       docReader,
	}
}

// tracker returns a usable SpanTracker — falls back to a no-op when the
// service was constructed without one.
func (s *ImageMultimodalService) tracker() SpanTracker {
	if s.spanTracker == nil {
		return noopSpanTracker{}
	}
	return s.spanTracker
}

// Handle implements asynq handler for TypeImageMultimodal.
func (s *ImageMultimodalService) Handle(ctx context.Context, task *asynq.Task) (retErr error) {
	var payload types.ImageMultimodalPayload
	if err := json.Unmarshal(task.Payload(), &payload); err != nil {
		return fmt.Errorf("unmarshal image multimodal payload: %w", err)
	}

	logger.Infof(ctx,
		"[ImageMultimodal] Processing image: chunk=%s, url=%s, attrs=%v",
		payload.ChunkID, payload.ImageURL, payload.ImageAttrsEnabled)

	ctx = context.WithValue(ctx, types.TenantIDContextKey, payload.TenantID)
	if payload.Language != "" {
		ctx = context.WithValue(ctx, types.LanguageContextKey, payload.Language)
	}

	// A reparse re-seeded the fan-in counter for its own images. Counting
	// this stale image would drain a slot of the new run and finalize it
	// before its images are done.
	if attemptSuperseded(ctx, s.tracker(), payload.KnowledgeID, payload.Attempt) {
		logger.Infof(ctx, "[ImageMultimodal] Attempt %d of %s superseded, dropping image %s",
			payload.Attempt, payload.KnowledgeID, payload.ImageURL)
		return nil
	}

	// Drop orphaned or user-aborted work before touching VLM. Missing
	// knowledge/KB rows are permanent failures — retrying only burns queue
	// capacity (asynq default MaxRetry=25 on legacy tasks).
	drop, dropErr := s.shouldDropOrphanedMultimodal(ctx, &payload)
	if dropErr != nil {
		if isFinalAsynqAttempt(ctx) {
			// No retry follows, so this image must still be counted or
			// the parent never reaches post-process.
			if err := s.checkAndFinalizeAllImages(ctx, payload); err != nil {
				return errors.Join(dropErr, err)
			}
		}
		return dropErr
	}
	if drop {
		logger.Infof(ctx,
			"[ImageMultimodal] Dropping task chunk=%s knowledge=%s kb=%s image=%s",
			payload.ChunkID, payload.KnowledgeID, payload.KnowledgeBaseID, payload.ImageURL)
		// Still count this image toward the parent finalize gate so a task
		// of dropped orphans cannot strand multimodal:pending forever.
		return s.checkAndFinalizeAllImages(ctx, payload)
	}

	tracker := s.tracker()
	// The trace map is shared with the per-image stage below: fields Handle
	// knows before the pipeline starts (which VLM ran) and fields the pipeline
	// discovers (attributes, policy, caption/OCR outcome) land on the same span.
	imgOut := types.JSONMap{}

	// finalize-once semantics: on success we always decrement the parent's
	// pending counter. On failure we only decrement when this is the last
	// asynq retry, so a permanently-failing image cannot leave the parent
	// knowledge stuck in "processing" forever — which was the #1 cause of
	// "stuck parsing" reports. Intermediate retries skip finalize so we don't
	// double-count and prematurely trigger post-process.
	var handleErr error
	defer func() {
		// Finalize on success, on the final asynq retry, or when the error is
		// a permanent failure wrapped with asynq.SkipRetry. The SkipRetry
		// branch is essential: a SkipRetry error makes the executor break
		// before the retry counter reaches maxRetry, so isFinalAsynqAttempt
		// would otherwise return false and the parent knowledge would be
		// stranded in "processing" forever.
		if handleErr == nil || isFinalAsynqAttempt(ctx) || errors.Is(handleErr, asynq.SkipRetry) {
			if err := s.checkAndFinalizeAllImages(ctx, payload); err != nil && retErr == nil {
				retErr = err
			}
		} else {
			logger.Infof(ctx,
				"[ImageMultimodal] Skip finalize on retryable error for %s (will count on last attempt)",
				payload.ImageURL)
		}
	}()
	vlmModel, vlmCfg, err := s.resolveVLM(ctx, payload.KnowledgeBaseID, payload.KnowledgeID)
	if err != nil {
		handleErr = fmt.Errorf("resolve VLM: %w", err)
		return handleErr
	}
	// Capture the resolved VLM model id (or "legacy_inline" for the legacy
	// inline-config path) so the trace shows WHICH model handled this image.
	// Without this, debugging "VLM is slow" requires a separate hop to the KB
	// config.
	if id := strings.TrimSpace(vlmCfg.ModelID); id != "" {
		imgOut["vlm_model_id"] = id
	} else {
		imgOut["vlm_model_id"] = "legacy_inline"
	}

	handleErr = wrapSkipRetry(s.processImage(ctx, &payload, vlmModel, vlmCfg, tracker, imgOut))
	return handleErr
}

// processImage runs one image pipeline for one image. The pipeline is chosen
// from the payload switches and owns the whole control flow; this function only
// hands it the run context and closes the span afterwards. The observation must
// be known before any OCR call is spent — that ordering is what makes "observe
// first, then act" possible: an image whose text is reliably absent does not pay
// for OCR.
//
// An observation (describe) failure is not fatal. The attributes then stay at
// their conservative defaults, which keeps OCR running, so a model that cannot
// observe costs one extra call rather than losing the text. This mirrors the
// upstream caption path, where a failed caption is also just a warning.
//
// out is the per-image trace map, owned by the caller and closed by the span
// this function ends; every field the pipeline discovers is written into it.
func (s *ImageMultimodalService) processImage(
	ctx context.Context,
	payload *types.ImageMultimodalPayload,
	vlmModel vlm.VLM,
	vlmCfg types.VLMConfig,
	tracker SpanTracker,
	out types.JSONMap,
) error {
	// Selected from the payload because a task in flight was enqueued for one
	// pipeline and must not be re-routed by an edit made since; the id also
	// lands on the trace, which reading image_info alone cannot do.
	pipeline := selectImagePipeline(payload)
	// Open a per-image subspan under the parent attempt's multimodal stage.
	// If the parent stage row is missing (legacy in-flight task, or the
	// upstream code shipped without span tracking), the tracker is a no-op so
	// we silently fall back to the existing counter-based finalize semantics.
	var imgSpan *Span
	if payload.Attempt > 0 {
		parent := tracker.LookupStage(ctx, payload.KnowledgeID, payload.Attempt, types.StageMultimodal)
		if parent != nil {
			name := fmt.Sprintf("multimodal.image[%d]", payload.ImageIndex)
			// The observation, and the policy it resolves to, are only known
			// after the describe round, so they land on the span's output (see
			// out[keyImageAttrs]) instead of here.
			imgSpan = tracker.BeginSubSpan(ctx, parent, name, types.SpanKindGeneration, types.JSONMap{
				"image_url":         payload.ImageURL,
				"image_source_type": payload.ImageSourceType,
				"parent_chunk_id":   payload.ChunkID,
				// Which pipeline ran. Reading image_info alone cannot tell a
				// default run from a smartocr run that observed and then
				// skipped OCR.
				"pipeline": string(pipeline.ID()),
			})
		}
	}

	var handleErr error
	defer func() {
		// Finalize the image subspan with the actual outcome — not the
		// finalize-counter outcome. The counter logic counts a "tried" image
		// regardless of inner success; the span surface tells the UI whether
		// THIS specific image worked.
		if imgSpan == nil {
			return
		}
		if handleErr == nil {
			tracker.EndSpan(ctx, imgSpan, out)
		} else {
			// Any genuinely failed attempt must close its span red, not
			// leave it dangling as "in progress". The span is a Generation
			// subspan, so FailSpan here cannot poison the parent stage or
			// the attempt status (cascade only fires for Stage-kind spans);
			// it merely records this attempt's outcome and lets a later
			// retry open its own span under the new attempt. A caller
			// cancellation gets its own code: the run did not fail, it was
			// walked away from.
			code := "MULTIMODAL_VLM_FAILED"
			if out["input_error"] != nil {
				code = "MULTIMODAL_INVALID_IMAGE"
			}
			if errors.Is(handleErr, context.Canceled) {
				code = "MULTIMODAL_CANCELLED"
			}
			tracker.FailSpan(ctx, imgSpan,
				code,
				handleErr.Error(),
				handleErr)
		}
	}()

	// Read image bytes. A provider:// URL must be resolved via FileService —
	// it must NEVER be handed to the HTTP downloader (which would fail with
	// "unsupported URL scheme"). On unrecoverable read failure the image is
	// skipped (the deferred finalize still counts it).
	imgBytes, readErr := s.readImageBytes(ctx, *payload)
	if readErr != nil {
		logger.Errorf(ctx, "[ImageMultimodal] Skip unreadable image %s: %v", payload.ImageURL, readErr)
		out[keyOutcome] = outcomeUnreadable
		out[keyReadError] = readErr.Error()
		handleErr = fmt.Errorf("read source image: %w", readErr)
		return handleErr
	}
	out[keyImageBytes] = len(imgBytes)
	// A successful storage read can still return an empty object or an HTML/
	// JSON error body. Do not let the VLM turn that into a successful caption
	// such as "please provide an image" and persist it in the gallery.
	if _, err := imageprep.Prepare(imgBytes, imageprep.Limits{}); err != nil {
		out["input_error"] = err.Error()
		handleErr = fmt.Errorf("%w: invalid source image: %v", asynq.SkipRetry, err)
		return handleErr
	}

	imageInfo := types.ImageInfo{
		SHA256:      fmt.Sprintf("%x", sha256.Sum256(imgBytes)),
		URL:         payload.ImageURL,
		OriginalURL: payload.ImageURL,
	}

	// The pipeline runs to completion, including the decision it makes about
	// OCR. Nothing below re-reads its result except the chunk building that
	// follows. The tracker and the image span travel with the run context so
	// the OCR action can raise and resolve its own subspan exactly like the
	// action-loop layout did (a failed OCR fails the .ocr subspan and the
	// outcome summary, without failing or retrying the whole image).
	rc := &runContext{
		payload:    payload,
		pipelineID: string(pipeline.ID()),
		params:     payload.ImagePipelineParams,
		declared:   pipeline.Fields(),
		model:      vlmModel,
		imageBytes: imgBytes,
		vlmCfg:     vlmCfg,
		imageInfo:  &imageInfo,
		out:        out,
		tracker:    tracker,
		imgSpan:    imgSpan,
		// The preprocess action asks this to straighten the page. Nil is fine:
		// the action reports "unavailable" and OCR reads the original bytes.
		preprocessor: s.docReader,
	}
	// Recorded here rather than inside a pipeline's Run: a run that returns early
	// with an error still did the actions it did, and "what ran before it failed"
	// is exactly what a trace is for.
	defer func() {
		if ids := rc.actionIDs(); len(ids) > 0 {
			out[keyActions] = ids
		}
	}()
	if err := pipeline.Run(ctx, rc); err != nil {
		// Surface the pipeline failure on this image's span instead of
		// letting the deferred finalize misread it as success. Without
		// this assignment the span below keeps handleErr == nil and the
		// UI shows a green (done) result for a genuinely failed image.
		handleErr = err
		return err
	}

	// Build child chunks for OCR and caption results
	imageInfoJSON, _ := json.Marshal([]types.ImageInfo{imageInfo})
	var newChunks []*types.Chunk

	if imageInfo.OCRText != "" {
		newChunks = append(newChunks, &types.Chunk{
			ID:              uuid.New().String(),
			TenantID:        payload.TenantID,
			KnowledgeID:     payload.KnowledgeID,
			KnowledgeBaseID: payload.KnowledgeBaseID,
			Content:         imageInfo.OCRText,
			ChunkType:       types.ChunkTypeImageOCR,
			ParentChunkID:   payload.ChunkID,
			SourceLocators:  payload.SourceLocators,
			IsEnabled:       true,
			Flags:           types.ChunkFlagRecommended,
			ImageInfo:       string(imageInfoJSON),
			CreatedAt:       time.Now(),
			UpdatedAt:       time.Now(),
		})
	}

	// A caption that reached the slot came from an action this run executed, so
	// an empty slot needs no second question asked of it.
	if imageInfo.Caption != "" {
		newChunks = append(newChunks, &types.Chunk{
			ID:              uuid.New().String(),
			TenantID:        payload.TenantID,
			KnowledgeID:     payload.KnowledgeID,
			KnowledgeBaseID: payload.KnowledgeBaseID,
			Content:         imageInfo.Caption,
			ChunkType:       types.ChunkTypeImageCaption,
			ParentChunkID:   payload.ChunkID,
			SourceLocators:  payload.SourceLocators,
			IsEnabled:       true,
			Flags:           types.ChunkFlagRecommended,
			ImageInfo:       string(imageInfoJSON),
			CreatedAt:       time.Now(),
			UpdatedAt:       time.Now(),
		})
	}
	out[keyChunksCreated] = len(newChunks)

	// Nothing was produced and a requested extraction failed: the image
	// has no image_info, so it will not appear in the gallery, and a silent
	// "skipped" success would hide a dead endpoint behind a green trace.
	// Report the failure to the task layer — the caller wraps permanent
	// failures (dead endpoint, bad key) with SkipRetry, transient ones burn
	// at most the asynq retry budget, and the deferred finalize marks this
	// image's span red (MULTIMODAL_CANCELLED when the run was walked away
	// from). The #4132 refinement: failures stay recorded-not-propagated only
	// while there is content to protect — a valid caption must not be
	// duplicated by an OCR retry; with nothing produced there is nothing to
	// protect, so the image fails honestly.
	extractionFailed := out[keyCaptionError] != nil || out[keyObservationError] != nil ||
		out[keyOCRResult] == "failed" ||
		out[keyOCRResult] == "skipped_observation_failed"
	if len(newChunks) == 0 && extractionFailed {
		out[keyOutcome] = outcomeFailed
		if ctx.Err() != nil {
			handleErr = fmt.Errorf("image processing cancelled with no content produced: %w", ctx.Err())
			return handleErr
		}
		if rc.vlmErr != nil {
			// The typed VLM error travels with the verdict so the task
			// layer can tell a permanent failure (SkipRetry) from a
			// transient one.
			handleErr = fmt.Errorf("image produced no content: %w", rc.vlmErr)
		} else {
			handleErr = fmt.Errorf("image produced no content (ocr_status=%v, caption_error=%v)",
				out[keyOCRResult], out[keyCaptionError])
		}
		return handleErr
	}

	if out[keyOCRResult] == "failed" {
		// Some content exists (a valid caption was persisted) — the failure
		// is recorded on the .ocr subspan and the outcome, and the image
		// stays a success at the task level (#4132).
		out[keyOutcome] = outcomePartialFailure
	}

	if len(newChunks) == 0 {
		// Deferred finalize will count this image on success. A genuine
		// blank (OCR answered no_text) or a policy-declined image keeps its
		// trace-level skip.
		out[keyOutcome] = outcomeNoContent
		return nil
	}

	// Persist chunks
	if err := s.chunkService.GetRepository().CreateChunks(ctx, newChunks); err != nil {
		handleErr = fmt.Errorf("create multimodal chunks: %w", err)
		return handleErr
	}
	for _, c := range newChunks {
		logger.Infof(ctx, "[ImageMultimodal] Created %s chunk %s for image %s, len=%d",
			c.ChunkType, c.ID, payload.ImageURL, len(c.Content))
	}

	// Index chunks so they can be retrieved
	s.indexChunks(ctx, *payload, newChunks, imgBytes, out)
	out[keyIndexed] = true
	// Only an image that reached this point with no verdict of its own is a
	// plain success: a partial failure was already recorded by the OCR clause
	// above, and overwriting it would hide the failure behind the content that
	// did survive.
	if _, decided := out[keyOutcome]; !decided {
		out[keyOutcome] = outcomeDone
	}

	return nil
}

// buildImageAttrsPrompt asks the model to observe the registered image
// attributes and describe the image in the same answer, so the attributes are
// known before the image's OCR call is spent. The attribute list and questions
// come from the registry, so the prompt cannot drift from the attributes the
// parser folds answers back onto.
//
// onDemand tells the builder whether the steps that consume on-demand
// attributes are switched on for this run. When they are off the prompt is
// byte-identical to what it was before those attributes existed, so a
// knowledge base that does not use them pays neither prompt length nor a
// chance of answer drift.
func buildImageAttrsPrompt(ctx context.Context, cfg types.VLMConfig, onDemand bool) string {
	language := strings.TrimSpace(cfg.DescriptionLanguage)
	if language == "" {
		language = types.LanguageNameFromContext(ctx)
	}
	var b strings.Builder
	fmt.Fprintf(&b,
		"Observe the following about the image, then describe its main content in %s.\n\n"+
			"Output one line per observation using exactly this format, then a DESCRIPTION line:\n\n",
		language)
	for _, spec := range types.ImageAttrRegistry {
		if !spec.WantsInjection(onDemand) {
			continue
		}
		allowed := make([]string, 0, len(spec.Values))
		for _, value := range spec.Values {
			allowed = append(allowed, value.Value)
		}
		fmt.Fprintf(&b, "%s: <%s>\n", spec.Name, strings.Join(allowed, " | "))
	}
	b.WriteString("DESCRIPTION: <one or two sentences>\n\nRules:\n")
	for _, spec := range types.ImageAttrRegistry {
		if !spec.WantsInjection(onDemand) {
			continue
		}
		fmt.Fprintf(&b, "- %s: %s\n", spec.Name, spec.Question)
	}
	b.WriteString("- Output only those lines. No preamble, no summary, no extra commentary.\n")
	return types.AppendCustomPromptInstructions(b.String(), cfg.CustomInstructions, "image_description")
}

// applyImageObservation copies a parsed verdict onto the image record so it is
// persisted with the chunk, and mirrors it into the trace map. The attributes
// are always kept — they are what the OCR policy reads next — while the
// description is stored only when this run wants a caption: the observe round is
// not optional even with captions off, because the attributes are the whole
// point of it.
func applyImageObservation(
	imageInfo *types.ImageInfo,
	attrs types.ImageAttrs,
	description string,
	out types.JSONMap,
	wantCaption bool,
) {
	if attrs.Attrs != nil {
		imageInfo.Attrs = attrs
	}
	if !wantCaption {
		out[keyObservationCaptionResult] = resultSkippedDisabled
		return
	}
	if description == "" {
		out[keyObservationCaptionResult] = "empty"
		return
	}
	imageInfo.Caption = description
	out[keyObservationCaptionChars] = len([]rune(description))
	out[keyObservationCaptionPreview] = previewText(description, 200)
	out[keyObservationCaptionResult] = "done"
}

// shouldDropOrphanedMultimodal reports whether the task should exit without
// retrying. True for user-cancelled/deleting knowledge, or when the parent
// knowledge / knowledge-base row no longer exists (deleted while queue entries
// survived).
func (s *ImageMultimodalService) shouldDropOrphanedMultimodal(
	ctx context.Context, payload *types.ImageMultimodalPayload,
) (bool, error) {
	if payload.KnowledgeID != "" && s.knowledgeRepo != nil {
		k, err := s.knowledgeRepo.GetKnowledgeByIDOnly(ctx, payload.KnowledgeID)
		if errors.Is(err, repository.ErrKnowledgeNotFound) {
			return true, nil
		}
		if err != nil {
			return false, err
		}
		switch k.ParseStatus {
		case types.ParseStatusCancelled, types.ParseStatusDeleting:
			return true, nil
		}
	}
	if payload.KnowledgeBaseID != "" && s.kbService != nil {
		kb, err := s.kbService.GetKnowledgeBaseByIDOnly(ctx, payload.KnowledgeBaseID)
		if errors.Is(err, repository.ErrKnowledgeBaseNotFound) {
			return true, nil
		}
		if err != nil {
			return false, err
		}
		if kb == nil {
			return true, nil
		}
	}
	return false, nil
}

// isFinalAsynqAttempt reports whether the current task context belongs to the
// last retry attempt before Asynq (or the Lite executor) archives the task. We use
// this to flip multimodal finalize semantics: during normal retries we skip
// counter decrement (the retry might still succeed), but on the final attempt
// we count the image regardless of outcome so a permanently-failing image
// cannot pin the parent knowledge in "processing" forever.
//
// Returns false when the values are unavailable (e.g. when the handler is
// invoked outside an asynq worker, as in unit tests). Treating that case as
// "not final" keeps test ergonomics — tests should drive finalize explicitly.
// isPermanentVLMFailure reports whether an error returned by the image run is
// a permanent failure that must NOT be retried by the task layer. Retrying
// such an error only burns queue capacity and re-hits the same dead end, so
// the caller wraps it with asynq.SkipRetry to archive the task immediately.
//
// Permanent (SkipRetry):
//   - a 4xx client error (bad key, oversized image, unsupported request): the
//     request cannot succeed on retry;
//   - a connection-level failure (connection refused / DNS / TLS / first-byte
//     timeout) surfaced as an api.TransportError in the "send request" phase:
//     the endpoint is unreachable, so every retry fails identically.
//
// Retryable (leave to the task-layer backoff):
//   - 5xx, 408, 429;
//   - a mid-stream transport error ("read response"): the stream broke after
//     the server answered, which is transient;
//   - anything else (unknown / context errors): safest to let the task retry.
//
// This is deliberately decoupled from the vlm package: it inspects only the
// stable api.* error types the manager always propagates. The manager wraps
// the original api error with %w, so errors.As still reaches it here.
func isPermanentVLMFailure(err error) bool {
	if err == nil {
		return false
	}
	var httpErr *api.HTTPError
	if errors.As(err, &httpErr) {
		switch {
		case httpErr.StatusCode == 429, // Too Many Requests
			httpErr.StatusCode == 408: // Request Timeout
			return false
		case httpErr.StatusCode >= 500:
			return false
		case httpErr.StatusCode >= 400:
			return true // permanent client error
		}
	}
	var transportErr *api.TransportError
	if errors.As(err, &transportErr) {
		// "send request" = failure before any response (connect / TLS /
		// first-byte timeout): endpoint is down. "read response" = the stream
		// broke mid-flight, which is transient.
		return transportErr.Op == "send request"
	}
	return false
}

// wrapSkipRetry marks a permanent VLM failure so the task layer archives the
// task immediately instead of retrying it. Retryable errors are returned
// unchanged.
func wrapSkipRetry(err error) error {
	if err == nil || !isPermanentVLMFailure(err) {
		return err
	}
	return fmt.Errorf("%w: %w", asynq.SkipRetry, err)
}

func isFinalAsynqAttempt(ctx context.Context) bool {
	retried, ok := asynq.GetRetryCount(ctx)
	if ok {
		maxRetry, maxRetryOK := asynq.GetMaxRetry(ctx)
		if maxRetryOK {
			return retried >= maxRetry
		}
	}
	retried, maxRetry, ok := types.TaskRetryMetadataFromContext(ctx)
	return ok && retried >= maxRetry
}

// indexChunks indexes the newly created multimodal chunks into the retrieval engine
// so they can participate in semantic search, then embeds the image itself
// when the knowledge base opted in and its embedding model takes images.
func (s *ImageMultimodalService) indexChunks(
	ctx context.Context, payload types.ImageMultimodalPayload, chunks []*types.Chunk,
	img []byte, out types.JSONMap,
) {
	kb, err := s.kbService.GetKnowledgeBaseByIDOnly(ctx, payload.KnowledgeBaseID)
	if err != nil || kb == nil {
		logger.Warnf(ctx, "[ImageMultimodal] Failed to get KB for indexing: %v", err)
		return
	}

	// Skip vector/keyword indexing when the KB has no embedding-based pipeline enabled
	// (e.g. Wiki-only KBs). Without this check, GetEmbeddingModel would fail because
	// EmbeddingModelID is intentionally empty for such KBs. The multimodal chunks
	// themselves are already persisted in the DB above, so skipping index here is safe.
	if !kb.NeedsEmbeddingModel() {
		logger.Infof(ctx, "[ImageMultimodal] Vector/keyword indexing disabled for KB %s, skipping index for %d multimodal chunks",
			kb.ID, len(chunks))
		// Still mark chunks as indexed so downstream finalization sees a consistent state.
		for _, chunk := range chunks {
			dbChunk, gerr := s.chunkService.GetChunkByIDOnly(ctx, chunk.ID)
			if gerr != nil {
				logger.Warnf(ctx, "[ImageMultimodal] Failed to fetch chunk %s for status update: %v", chunk.ID, gerr)
				continue
			}
			dbChunk.Status = int(types.ChunkStatusIndexed)
			if uerr := s.chunkService.GetRepository().UpdateChunk(ctx, dbChunk); uerr != nil {
				logger.Warnf(ctx, "[ImageMultimodal] Failed to update chunk %s status to indexed: %v", chunk.ID, uerr)
			}
		}
		return
	}

	embeddingModel, err := s.modelService.GetEmbeddingModel(ctx, kb.EmbeddingModelID)
	if err != nil {
		logger.Warnf(ctx, "[ImageMultimodal] Failed to get embedding model for indexing: %v", err)
		return
	}

	tenantInfo, err := s.tenantRepo.GetTenantByID(ctx, payload.TenantID)
	if err != nil {
		logger.Warnf(ctx, "[ImageMultimodal] Failed to get tenant for indexing: %v", err)
		return
	}
	// The factory's unbound path reads TenantInfo from ctx; make sure it's there.
	ctx = context.WithValue(ctx, types.TenantInfoContextKey, tenantInfo)

	// Resolve engine via the factory using the KB's VectorStore binding
	// (nil → tenant effective engines fallback; verified tenant ownership otherwise).
	engine, err := retriever.CreateRetrieveEngineForKB(
		ctx, s.retrieveEngine, s.ownership, payload.TenantID, kb.VectorStoreID)
	if err != nil {
		logger.Warnf(ctx, "[ImageMultimodal] Failed to init retrieve engine: %v", err)
		return
	}

	indexInfoList := make([]*types.IndexInfo, 0, len(chunks))
	for _, chunk := range chunks {
		indexInfoList = append(indexInfoList, &types.IndexInfo{
			Content:         chunk.Content,
			SourceID:        chunk.ID,
			SourceType:      types.ChunkSourceType,
			ChunkID:         chunk.ID,
			KnowledgeID:     chunk.KnowledgeID,
			KnowledgeBaseID: chunk.KnowledgeBaseID,
			// Engines other than Postgres store the zero value verbatim and
			// filter on is_enabled = true, so leaving this unset made OCR and
			// caption chunks unsearchable there.
			IsEnabled: chunk.IsEnabled,
		})
	}

	if err := engine.BatchIndex(ctx, embeddingModel, indexInfoList); err != nil {
		logger.Errorf(ctx, "[ImageMultimodal] Failed to index multimodal chunks: %v", err)
		return
	}

	// Mark chunks as indexed.
	// Must re-fetch from DB because the in-memory objects lack auto-generated fields
	// (e.g. seq_id), and GORM Save would overwrite them with zero values.
	for _, chunk := range chunks {
		dbChunk, err := s.chunkService.GetChunkByIDOnly(ctx, chunk.ID)
		if err != nil {
			logger.Warnf(ctx, "[ImageMultimodal] Failed to fetch chunk %s for status update: %v", chunk.ID, err)
			continue
		}
		dbChunk.Status = int(types.ChunkStatusIndexed)
		if err := s.chunkService.GetRepository().UpdateChunk(ctx, dbChunk); err != nil {
			logger.Warnf(ctx, "[ImageMultimodal] Failed to update chunk %s status to indexed: %v", chunk.ID, err)
		}
	}

	logger.Infof(ctx, "[ImageMultimodal] Indexed %d multimodal chunks for knowledge %s",
		len(chunks), payload.KnowledgeID)

	if status := s.indexImageVector(ctx, kb, payload, img, chunks, embeddingModel, engine); status != "" {
		out["image_vector"] = status
	}
}

// indexImageVector embeds the image itself with a multimodal embedding model
// and stores the vector under an image_vector chunk, so a query can find the
// image by what it shows even where the caption left that out. It returns a
// short status for the trace, "" when image indexing is disabled for this
// document (falling back to the KB setting) or the model takes no images.
//
// The caption and OCR chunks are already indexed, so a failure here only
// loses the extra recall and never fails the task.
func (s *ImageMultimodalService) indexImageVector(
	ctx context.Context, kb *types.KnowledgeBase, payload types.ImageMultimodalPayload, img []byte,
	chunks []*types.Chunk, model embedding.Embedder, engine *retriever.CompositeRetrieveEngine,
) string {
	// The task's explicit choice wins over the KB default, including false.
	// A vector index is still required; old tasks inherit the KB setting.
	enabled := kb.IsImageVectorEnabled()
	if payload.ImageVectorEnabled != nil {
		enabled = kb.IsVectorEnabled() && *payload.ImageVectorEnabled
	}
	if !enabled {
		return ""
	}
	imageModel, ok := embedding.AsImageEmbedder(model)
	if !ok {
		return ""
	}
	// A scanned page is text, which its OCR chunk already carries better
	// than an image vector can: multimodal embeddings are weakest on dense
	// text, and a document of such pages would cost a call each.
	if payload.ImageSourceType == "scanned_pdf" {
		return "skipped: scanned page"
	}
	// The chunk needs text for reranking and the answer context; the caption
	// describes the image, the OCR text is the fallback.
	var source *types.Chunk
	for _, c := range chunks {
		if c.ChunkType == types.ChunkTypeImageCaption {
			source = c
			break
		}
		if c.ChunkType == types.ChunkTypeImageOCR {
			source = c
		}
	}
	if source == nil {
		return "skipped: no caption or OCR text"
	}

	prepared, err := imageprep.Prepare(img, imageModel.ImageLimits())
	if err != nil {
		logger.Warnf(ctx, "[ImageMultimodal] Image %s not embeddable: %v", payload.ImageURL, err)
		return "failed: " + err.Error()
	}
	vectors, err := imageModel.BatchEmbedImages(ctx, []embedding.Image{prepared})
	if err != nil {
		logger.Warnf(ctx, "[ImageMultimodal] Image embedding failed for %s: %v", payload.ImageURL, err)
		return "failed: " + err.Error()
	}
	if dims := model.GetDimensions(); dims > 0 && len(vectors[0]) != dims {
		// The index is laid out for the text vectors' width; a vector of
		// another would land in a different collection or be refused.
		logger.Warnf(ctx, "[ImageMultimodal] Image vector for %s has %d dimensions, the index %d",
			payload.ImageURL, len(vectors[0]), dims)
		return fmt.Sprintf("failed: %d dimensions, index has %d", len(vectors[0]), dims)
	}

	now := time.Now()
	chunk := &types.Chunk{
		ID:              uuid.New().String(),
		TenantID:        payload.TenantID,
		KnowledgeID:     payload.KnowledgeID,
		KnowledgeBaseID: payload.KnowledgeBaseID,
		Content:         source.Content,
		ChunkType:       types.ChunkTypeImageVector,
		ParentChunkID:   payload.ChunkID,
		IsEnabled:       true,
		Flags:           types.ChunkFlagRecommended,
		ImageInfo:       source.ImageInfo,
		CreatedAt:       now,
		UpdatedAt:       now,
	}
	if err := s.chunkService.GetRepository().CreateChunks(ctx, []*types.Chunk{chunk}); err != nil {
		logger.Warnf(ctx, "[ImageMultimodal] Failed to create image vector chunk for %s: %v", payload.ImageURL, err)
		return "failed: " + err.Error()
	}
	err = engine.BatchIndexVectors(ctx, model, []*types.IndexInfo{{
		Content:         chunk.Content,
		SourceID:        chunk.ID,
		SourceType:      types.ImageSourceType,
		ChunkID:         chunk.ID,
		KnowledgeID:     chunk.KnowledgeID,
		KnowledgeBaseID: chunk.KnowledgeBaseID,
		IsEnabled:       true,
	}}, vectors)
	if err != nil {
		// The chunk stays unindexed: nothing retrieves it, and a re-parse
		// replaces it with the rest of the knowledge's chunks.
		logger.Warnf(ctx, "[ImageMultimodal] Failed to index image vector for %s: %v", payload.ImageURL, err)
		return "failed: " + err.Error()
	}
	// Re-fetch for the same reason as indexChunks: GORM Save would zero
	// the generated fields the in-memory chunk lacks.
	if dbChunk, err := s.chunkService.GetChunkByIDOnly(ctx, chunk.ID); err == nil {
		dbChunk.Status = int(types.ChunkStatusIndexed)
		if err := s.chunkService.GetRepository().UpdateChunk(ctx, dbChunk); err != nil {
			logger.Warnf(ctx, "[ImageMultimodal] Failed to mark image vector chunk %s indexed: %v", chunk.ID, err)
		}
	}
	logger.Infof(ctx, "[ImageMultimodal] Indexed image vector chunk %s for image %s", chunk.ID, payload.ImageURL)
	return "indexed"
}

// resolveVLM creates a vlm.VLM instance for the given knowledge base,
// supporting both new-style (ModelID) and legacy (inline BaseURL) configs.
// Per-upload process_overrides on the knowledge entry take precedence over KB defaults.
func (s *ImageMultimodalService) resolveVLM(ctx context.Context, kbID, knowledgeID string) (vlm.VLM, types.VLMConfig, error) {
	kb, err := s.kbService.GetKnowledgeBaseByIDOnly(ctx, kbID)
	if err != nil {
		return nil, types.VLMConfig{}, fmt.Errorf("get knowledge base %s: %w", kbID, err)
	}
	if kb == nil {
		return nil, types.VLMConfig{}, fmt.Errorf("knowledge base %s not found", kbID)
	}

	var processOverrides *types.KnowledgeProcessOverrides
	if knowledgeID != "" && s.knowledgeRepo != nil {
		if k, kerr := s.knowledgeRepo.GetKnowledgeByIDOnly(ctx, knowledgeID); kerr == nil && k != nil {
			processOverrides, _ = k.ProcessOverrides()
		}
	}
	vlmCfg := ResolveProcessConfig(kb, processOverrides).VLMConfig
	if !vlmCfg.IsEnabled() {
		return nil, types.VLMConfig{}, fmt.Errorf("VLM is not enabled for knowledge base %s", kbID)
	}

	// New-style: resolve model through ModelService
	if vlmCfg.ModelID != "" {
		model, err := s.modelService.GetVLMModel(ctx, vlmCfg.ModelID)
		if err == nil {
			return model, vlmCfg, nil
		}

		// A shared KB is processed in its owner's tenant context, while a
		// document uploaded by an editor may carry a model ID from the
		// editor's tenant. If that document-level model cannot be resolved,
		// retry with the KB owner's configured model instead.
		kbVLMConfig := kb.VLMConfig
		if vlmCfg.ModelID != kbVLMConfig.ModelID && kbVLMConfig.ModelID != "" && kbVLMConfig.IsEnabled() {
			logger.Warnf(ctx,
				"[ImageMultimodal] Failed to resolve document VLM model %s: %v;"+
					" falling back to knowledge base VLM model %s",
				vlmCfg.ModelID, err, kbVLMConfig.ModelID)
			fallbackModel, fallbackErr := s.modelService.GetVLMModel(ctx, kbVLMConfig.ModelID)
			if fallbackErr != nil {
				logger.Warnf(ctx,
					"[ImageMultimodal] Failed to resolve knowledge base fallback VLM model %s: %v",
					kbVLMConfig.ModelID, fallbackErr)
				return nil, vlmCfg, errors.Join(
					fmt.Errorf("resolve document VLM model %s: %w", vlmCfg.ModelID, err),
					fmt.Errorf("resolve knowledge base fallback VLM model %s: %w", kbVLMConfig.ModelID, fallbackErr),
				)
			}
			return fallbackModel, kbVLMConfig, nil
		}
		return nil, vlmCfg, err
	}

	// Legacy: create VLM from inline config
	model, err := vlm.NewVLMFromLegacyConfig(vlmCfg, s.ollamaService)
	return model, vlmCfg, err
}

// resolveFileServiceForPayload resolves tenant/KB scoped file service for reading provider:// URLs.
// Falls back to the globally configured default FileService when the tenant's
// StorageEngineConfig does not carry a usable configuration for the URL's provider.
// This mirrors the write-side fallback in knowledgeService.resolveFileService
// and is required because images can be saved using global STORAGE_TYPE/MINIO_*
// env vars while tenant.StorageEngineConfig.MinIO is left empty (issue #1282).
func (s *ImageMultimodalService) resolveFileServiceForPayload(
	ctx context.Context, payload types.ImageMultimodalPayload,
) interfaces.FileService {
	kb := func() *types.KnowledgeBase {
		kb, err := s.kbService.GetKnowledgeBaseByIDOnly(ctx, payload.KnowledgeBaseID)
		if err != nil {
			logger.Warnf(ctx, "[ImageMultimodal] GetKnowledgeBaseByIDOnly failed: kb=%s err=%v",
				payload.KnowledgeBaseID, err)
			return nil
		}
		return kb
	}
	return s.imageReader().fileService(ctx, payload.TenantID, kb, payload.ImageURL)
}

func (s *ImageMultimodalService) imageReader() knowledgeImageReader {
	return knowledgeImageReader{
		tenantRepo:      s.tenantRepo,
		fileSvc:         s.fileSvc,
		storageResolver: s.storageResolver,
		resourceCatalog: s.resourceCatalog,
	}
}

// readImageBytes loads the image bytes for one image of a multimodal payload.
//   - For provider:// URLs (local://, minio://, s3://, cos://, ...) it reads via
//     the resolved FileService and NEVER falls back to HTTP — handing a
//     provider:// URL to the HTTP downloader is what caused issue #1282.
//   - For legacy in-flight payloads with ImageLocalPath set, it tries the local
//     file before falling back to the URL.
//   - For plain http(s):// URLs it uses the SSRF-safe downloader.
//
// imageURL is passed explicitly (instead of read from payload.ImageURL) because
// a batched payload carries several images and only the first one is mirrored
// onto the legacy single-image fields.
func (s *ImageMultimodalService) readImageBytes(ctx context.Context, payload types.ImageMultimodalPayload) ([]byte, error) {
	if isStoredImageURL(payload.ImageURL) {
		fileSvc := s.resolveFileServiceForPayload(ctx, payload)
		if fileSvc == nil {
			return nil, fmt.Errorf("no file service available for %s", payload.ImageURL)
		}
		reader, err := fileSvc.GetFile(ctx, payload.ImageURL)
		if err != nil {
			return nil, fmt.Errorf("file service get %s: %w", payload.ImageURL, err)
		}
		defer reader.Close()
		data, err := io.ReadAll(reader)
		if err != nil {
			return nil, fmt.Errorf("read %s: %w", payload.ImageURL, err)
		}
		return data, nil
	}

	if payload.ImageLocalPath != "" {
		if data, err := os.ReadFile(payload.ImageLocalPath); err == nil {
			return data, nil
		} else {
			logger.Warnf(ctx, "[ImageMultimodal] Local file %s not available (%v), falling back to URL", payload.ImageLocalPath, err)
		}
	}

	data, err := downloadImageFromURL(payload.ImageURL)
	if err != nil {
		return nil, fmt.Errorf("download %s: %w", payload.ImageURL, err)
	}
	logger.Infof(ctx, "[ImageMultimodal] Image downloaded from URL, len=%d", len(data))
	return data, nil
}

// downloadImageFromURL downloads image bytes from an HTTP(S) URL.
func downloadImageFromURL(imageURL string) ([]byte, error) {
	return secutils.DownloadBytes(imageURL)
}

// multimodalPendingKey is the Redis key holding how many image tasks a
// knowledge is still waiting on. The fan-out (enqueueImageMultimodalTasks)
// seeds it and the fan-in (checkAndFinalizeAllImages) drains it, so the two
// must agree on the name — hence one helper rather than two format strings.
func multimodalPendingKey(knowledgeID string) string {
	return fmt.Sprintf("multimodal:pending:%s", knowledgeID)
}

// checkAndFinalizeAllImages counts one image of the parent knowledge and,
// once every image is counted, hands the knowledge to post-process. An error
// means post-process could not be enqueued; the caller must fail the task so
// the row is not left in "processing" with nobody to move it on.
func (s *ImageMultimodalService) checkAndFinalizeAllImages(
	ctx context.Context, payload types.ImageMultimodalPayload,
) error {
	redisKey := multimodalPendingKey(payload.KnowledgeID)

	var pendingCount int64
	if s.redisClient == nil {
		pendingCount = liteMultimodalDecrBy(redisKey, 1)
	} else {
		var err error
		pendingCount, err = s.redisClient.Decr(ctx, redisKey).Result()
		if err != nil && err != redis.Nil {
			// Redis hiccup must not strand the parent knowledge. Best-effort:
			// enqueue post-process anyway. KnowledgePostProcess is idempotent
			// (it transitions parse_status processing → completed under a row
			// guard), so a duplicate triggered by a sibling image is harmless.
			// The alternative — silently returning — is what produced the
			// "permanently stuck" reports we are fixing here.
			logger.Warnf(ctx,
				"[ImageMultimodal] Decrement failed for %s (%v); fallback-enqueueing post-process",
				payload.KnowledgeID, err)
			return s.enqueueKnowledgePostProcessTask(ctx, payload)
		}
	}

	if pendingCount > 0 {
		return nil
	}
	logger.Infof(ctx, "[ImageMultimodal] All images processed for knowledge %s. Finalizing...", payload.KnowledgeID)
	// Enqueue before dropping the counter: a retry of this task then finds
	// the key at or below zero and enqueues again.
	if err := s.enqueueKnowledgePostProcessTask(ctx, payload); err != nil {
		return err
	}
	if s.redisClient == nil {
		liteMultimodalDel(redisKey)
	} else {
		s.redisClient.Del(ctx, redisKey)
	}
	return nil
}

func (s *ImageMultimodalService) enqueueKnowledgePostProcessTask(
	ctx context.Context, payload types.ImageMultimodalPayload,
) error {
	if s.taskEnqueuer == nil {
		return nil
	}

	taskPayload := types.KnowledgePostProcessPayload{
		TenantID:        payload.TenantID,
		KnowledgeID:     payload.KnowledgeID,
		KnowledgeBaseID: payload.KnowledgeBaseID,
		Language:        payload.Language,
		// Lets post-process skip itself when a reparse superseded the run.
		Attempt: payload.Attempt,
	}
	langfuse.InjectTracing(ctx, &taskPayload)
	payloadBytes, err := json.Marshal(taskPayload)
	if err != nil {
		logger.Warnf(ctx, "[ImageMultimodal] Failed to marshal post process payload: %v", err)
		return fmt.Errorf("marshal post process payload: %w", err)
	}

	task := asynq.NewTask(types.TypeKnowledgePostProcess, payloadBytes,
		knowledgePostProcessTaskOptions()...)
	if err := enqueueWithRetry(ctx, s.taskEnqueuer, task); err != nil {
		logger.Errorf(ctx, "[ImageMultimodal] Failed to enqueue post process task for %s: %v", payload.KnowledgeID, err)
		return err
	}
	logger.Infof(ctx, "[ImageMultimodal] Enqueued post process task for %s", payload.KnowledgeID)
	return nil
}