Skip to content

RAGFlow is an agent-native database: a Retrieval-Augmented Generation (RAG) engine that grew into context infrastructure.

28.3 KB
Raw Blame
//
//  Copyright 2026 The InfiniFlow Authors. All Rights Reserved.
//
//  Licensed under the Apache License, Version 2.0 (the "License");
//  you may not use this file except in compliance with the License.
//  You may obtain a copy of the License at
//
//      http://www.apache.org/licenses/LICENSE-2.0
//
//  Unless required by applicable law or agreed to in writing, software
//  distributed under the License is distributed on an "AS IS" BASIS,
//  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
//  See the License for the specific language governing permissions and
//  limitations under the License.
//

// Package component — Parser component (Phase 2.2 of
// port-rag-flow-pipeline-to-go.md §4).
//
// SCOPE (honest):
//
//   - WHAT IS PORTED:
//
//   - The component's lifecycle contract: NewParserComponent /
//     Invoke / Inputs / Outputs and registration under
//     runtime.CategoryIngestion.
//
//   - Per-page parallelism is delegated to the parser backends
//     (e.g. internal/deepdoc/parser/pdf fans out one worker per
//     page and assembles the results in page order). This
//     component normalizes the parser output into structured JSON
//     items while preserving the backend's deterministic item order.
//
//   - Progress (start/done callback) and elapsed-time stamping
//     (_created_time / _elapsed_time) are owned by the canvas
//     framework (internal/agent/canvas/node_body.go realComponentBody),
//     which wraps every component Invoke. This component does not call
//     those helpers itself. See internal/agent/runtime/helpers.go.
//
//   - WHAT IS NOT YET PORTED:
//
//   - The Python component dispatches to 13 file-format branches
//     (pdf, Markdown, text&code, html, spreadsheet, slides, doc,
//     docx, image, audio, video, email, epub) — see parser.py
//     function_map at line ~1273. The Go port is LANDed and LIVE
//     in production for the families the ingestor claims (see
//     cmd/ragflow_server.go: Ingestor.supportedTypes =
//     ["pdf","docx","txt"]); those run their real parsers,
//     including the cgo-gated office variants via office_oxide.
//     Families not yet ported fall through to the raw-text path
//     below rather than printing skeletons.
//
//   - For any family NOT yet ported (its Go parser returns no real
//     data), the component uses a "raw text" fallback: it treats
//     the input binary as UTF-8 and slices it into 1 page (or N
//     pages when the upstream signals a page boundary with a literal
//     "\f" form feed). This is the conservative, observable
//     behaviour for UNPORTED families only; ported families run
//     their real parsers.
//
//   - The Python side's "image2id" pipeline (parser.py:1317-1329)
//     that uploads embedded images to MinIO is not replicated —
//     the schema layer carries images as opaque map values, and
//     the upload step is the responsibility of a separate
//     side-effect component (out of scope for Phase 2.2).
//
//   - The Python _param.check() business validation (parse_method whitelist)
//     is mirrored by (*ParserComponent).Check() below, which NewParserComponent
//     runs at construction time. Neither backend validates
//     audio/video vlm.llm_id: Python's check() has no such branch,
//     and the audio model is resolved at dispatch time with a
//     tenant-default fallback.
//
//   - NO PERSISTENCE: structured parser items live only in the per-run
//     output map.
package component

import (
	"context"
	"errors"
	"fmt"
	"strings"
	"unicode/utf8"

	"go.uber.org/zap"
	"gorm.io/gorm"

	"ragflow/internal/agent/runtime"
	"ragflow/internal/common"
	"ragflow/internal/ingestion/component/globals"
	"ragflow/internal/ingestion/component/progress"
	"ragflow/internal/ingestion/component/schema"
	"ragflow/internal/parser/parser"
	"ragflow/internal/service/billing"
	"ragflow/internal/utility"
)

// ComponentNameParser is the canonical Parser component name; the value lives
// in common and is re-exported by the schema package so both layers reference
// the same identity.
const ComponentNameParser = schema.ComponentNameParser

// pageFormFeed is the byte that text-page mode treats as a
// hard page boundary. Matches the ASCII form feed (\f, 0x0C) — the
// same convention used by the Python TxtParser and by most
// "page-segmented text" codecs.
const pageFormFeed = '\f'

// ParserComponent runs the configured parser branch against the
// upstream "binary" payload and returns structured parser outputs.
//
// The instance is safe for concurrent invocation: each Invoke call
// builds its own per-batch goroutine tree and merges results in
// the goroutine that returned from Invoke. The static Param is
// read-only after construction.
type ParserComponent struct {
	setups                  map[string]schema.ParserSetup
	enableVisionEnhancement bool
	vision                  visionSettings
}

// visionSettings carries the global vision-enhancement choices. They sit at the
// Parser params top level (vlm) instead of a per-family setup, so one model and
// prompt serve the enhanced file types rather than each family repeating them.
// Response language is deliberately absent: it belongs to the knowledge base,
// which also drives tokenization, so the two can never disagree.
type visionSettings struct {
	modelID      string
	systemPrompt string
}

// NewParserComponent constructs a Parser from a DSL param map.
// The default setups are overlaid with the supplied values. The Parser
// emits one output protocol: structured JSON items for every family,
// including PDF and office documents. Markdown remains an internal backend
// representation only and is never a public Parser output.
//
// Param map shape (all keys optional):
//
//	{
//	  "enable_vision_enhancement": bool,
//	  "vlm":                  {"llm_id": string, "system_prompt": string},
//	  "pdf":                  map[string]any,
//	  "docx":                 map[string]any,
//	  ...
//	}
//
// Errors here surface as canvas compile failures so a malformed
// param is caught at build time rather than mid-run.
func NewParserComponent(params map[string]any) (runtime.Component, error) {
	s := defaultSetups()
	if params == nil {
		return &ParserComponent{setups: s}, nil
	}
	// Canvases saved by the Python-era frontend nest the per-family setups
	// under a "setups" key; lift them so every family lands at the top level.
	params = schema.FlattenLegacyParserSetups(params)
	var enableVisionEnhancement bool
	if raw, exists := params["enable_vision_enhancement"]; exists {
		var ok bool
		enableVisionEnhancement, ok = raw.(bool)
		if !ok {
			return nil, errors.New("parser: enable_vision_enhancement must be a boolean")
		}
	}
	var vision visionSettings
	if raw, exists := params["vlm"]; exists {
		vlm, ok := raw.(map[string]any)
		if !ok {
			return nil, errors.New("parser: vlm must be an object")
		}
		if rawID, exists := vlm["llm_id"]; exists {
			vision.modelID, ok = rawID.(string)
			if !ok {
				return nil, errors.New("parser: vlm.llm_id must be a string")
			}
		}
		if rawPrompt, exists := vlm["system_prompt"]; exists {
			vision.systemPrompt, ok = rawPrompt.(string)
			if !ok {
				return nil, errors.New("parser: vlm.system_prompt must be a string")
			}
		}
	}
	for k, raw := range params {
		if k == "outputs" || k == "allowed_output_format" || k == "enable_vision_enhancement" || k == "vlm" {
			continue
		}
		ftCfg, ok := raw.(map[string]any)
		if !ok {
			continue
		}
		if _, exists := s[k]; !exists {
			s[k] = schema.ParserSetup{}
		}
		for fk, fv := range ftCfg {
			s[k][fk] = cloneParserSetupValue(fv)
		}
	}
	pc := &ParserComponent{setups: s, enableVisionEnhancement: enableVisionEnhancement, vision: vision}
	if err := pc.Check(); err != nil {
		return nil, fmt.Errorf("parser: %w", err)
	}
	return pc, nil
}

func cloneParserSetupValue(value any) any {
	switch v := value.(type) {
	case map[string]any:
		cloned := make(map[string]any, len(v))
		for key, nested := range v {
			cloned[key] = cloneParserSetupValue(nested)
		}
		return cloned
	case schema.ParserSetup:
		cloned := make(schema.ParserSetup, len(v))
		for key, nested := range v {
			cloned[key] = cloneParserSetupValue(nested)
		}
		return cloned
	case []any:
		cloned := make([]any, len(v))
		for i, nested := range v {
			cloned[i] = cloneParserSetupValue(nested)
		}
		return cloned
	case []string:
		return append([]string(nil), v...)
	case []int:
		return append([]int(nil), v...)
	case [][]int:
		cloned := make([][]int, len(v))
		for i, nested := range v {
			cloned[i] = append([]int(nil), nested...)
		}
		return cloned
	default:
		return value
	}
}

// Check validates parser methods at construction time so a malformed DSL
// surfaces as a canvas compile failure rather than a mid-run error.
//
// NOT covered here (intentional):
//   - audio/video vlm.llm_id: Python's check() does not validate it
//     either, and audio dispatch resolves a missing/empty model to
//     the tenant default, so validating it here would only block
//     otherwise valid pipelines (see ingestion_pipeline_audio.json).
//   - any language: a missing language is no longer fatal downstream —
//     captions fall back to the knowledge base language and MinerU to its own
//     default — and Check runs before run inputs exist, so the knowledge base
//     value cannot be consulted here. Rejecting stored canvases for it would
//     only break pipelines that resolve fine at run time.
func (c *ParserComponent) Check() error {
	if pdf, ok := c.setups["pdf"]; ok {
		if pm, _ := pdf["parse_method"].(string); pm == "" {
			return errors.New("parse method abnormal. does not support empty value")
		}
		if _, err := readMinerUExtractionOptions(pdf); err != nil {
			return err
		}
	}
	// The whole image contract keys off whether ocr_enabled is present, so a
	// value of the wrong type must not quietly read as "absent" and fall back to
	// the legacy parse_method inference.
	if img, ok := c.setups["image"]; ok {
		if raw, exists := img["ocr_enabled"]; exists {
			if _, isBool := raw.(bool); !isBool {
				return errors.New("parser: image ocr_enabled must be a boolean")
			}
		}
		if raw, exists := img["ocr_model_id"]; exists {
			if _, isString := raw.(string); !isString {
				return errors.New("parser: image ocr_model_id must be a string")
			}
		}
		// An image needs at least one text source: local OCR or the global
		// vision enhancement. With neither, the item carries no text and the
		// Tokenizer drops it, so reject the configuration at build time
		// instead of indexing nothing silently.
		if !c.enableVisionEnhancement && !imageSetupRunsOCR(img) {
			return errors.New("parser: image has no text source: OCR is off and vision enhancement is disabled")
		}
	}
	return nil
}

// imageSetupRunsOCR mirrors the OCR selection in maybeDispatchImage
// (media_dispatch.go): the ocr_enabled switch wins when present, otherwise
// the legacy parse_method reading applies ("ocr"/empty runs local OCR).

func imageSetupRunsOCR(setup schema.ParserSetup) bool {
	if ocrEnabled, hasOCRFlag := setup["ocr_enabled"].(bool); hasOCRFlag {
		return ocrEnabled
	}
	method := getStringOr(setup, "parse_method", "")
	return method == "" || strings.EqualFold(method, "ocr")
}

func defaultSetups() map[string]schema.ParserSetup {
	s := map[string]schema.ParserSetup{
		"pdf": {
			"parse_method":          "deepdoc",
			"flatten_media_to_text": false,
			"remove_toc":            false,
			"remove_header_footer":  false,
			"suffix":                []string{"pdf"},
		},
		"spreadsheet": {
			"parse_method":          "deepdoc",
			"flatten_media_to_text": false,
			"html4excel":            false,
			"suffix":                []string{"xls", "xlsx", "csv"},
		},
		"doc": {
			"remove_toc":           false,
			"remove_header_footer": false,
			"suffix":               []string{"doc"},
		},
		"docx": {
			"flatten_media_to_text": false,
			"remove_toc":            false,
			"remove_header_footer":  false,
			"suffix":                []string{"docx"},
		},
		"markdown": {
			"flatten_media_to_text": false,
			"suffix":                []string{"md", "markdown", "mdx"},
			"remove_toc":            false,
		},
		"text&code": {
			"suffix": []string{
				"txt", "py", "js", "java", "c", "cpp", "h", "php",
				"go", "ts", "sh", "cs", "kt", "sql",
			},
		},
		"html": {
			"suffix":               []string{"htm", "html"},
			"remove_toc":           false,
			"remove_header_footer": false,
		},
		"slides": {
			"parse_method": "deepdoc",
			"suffix":       []string{"pptx", "ppt"},
		},
		"image": {
			// Default stays on the legacy parse_method shape. ocr_enabled is
			// deliberately NOT set here: NewParserComponent overlays params on
			// top of these defaults, so the mere presence of ocr_enabled in the
			// merged setup proves the caller chose the switch shape. A default
			// value would leak the flag onto legacy (parse_method-only) canvases.
			"parse_method":  "ocr",
			"llm_id":        "",
			"system_prompt": "",
			"suffix":        []string{"jpg", "jpeg", "png", "gif", "bmp", "tif", "tiff", "webp"},
		},
		"email": {
			"suffix": []string{"eml", "msg"},
			"fields": []string{
				"from", "to", "cc", "bcc", "date", "subject",
				"body", "attachments", "metadata",
			},
		},
		"audio": {
			"suffix": []string{
				"da", "wave", "wav", "mp3", "aac", "flac", "ogg",
				"aiff", "au", "midi", "wma", "realaudio", "vqf",
				"oggvorbis", "ape",
			},
		},
		"video": {
			"suffix": []string{"mp4", "avi", "mkv"},
			"prompt": "",
		},
		"epub": {
			"suffix": []string{"epub"},
		},
		"json": {
			"suffix": []string{"json", "jsonl", "ldjson"},
		},
	}
	return s
}

// Inputs returns the static parameter metadata. The component
// reads the following from the inputs map at Invoke time:
//
//	binary    ([]byte, optional) — file bytes from upstream File.
//	                                When absent, Parser resolves them from
//	                                bucket/path or doc_id.
//	name      (string, optional) — resolved source filename.
//	file      (map[string]any, optional) — source descriptor; its name is
//	                                used when name is absent.
//	file_type (string, optional) — explicit parser routing hint.
//	lang      (string, optional) — language forwarded to downstream stages.
//	doc_id    (string, optional) — document ID used for naming and,
//	                                when binary is absent, storage lookup.
func (c *ParserComponent) Inputs() map[string]string {
	return map[string]string{
		"binary":    "Optional file bytes ([]byte). When absent, Parser resolves them from bucket/path or doc_id.",
		"name":      "Optional resolved source filename. Takes precedence over file.name.",
		"file":      "Optional source file descriptor (map[string]any). file.name is used when name is absent.",
		"file_type": "Optional explicit parser routing hint (string).",
		"lang":      "Optional language for downstream tokenization (string).",
		"doc_id":    "Optional document ID (string). Used for downstream correlation and doc_id-driven storage lookup.",
		"bucket":    "Optional storage bucket override. Used when binary is absent.",
		"path":      "Optional storage object key override. Used when binary is absent.",
	}
}

// Outputs returns the public surface that downstream ingestion
// components (Chunker, Tokenizer, Extractor) can wire into.
//
//	name          string  — carried over from the upstream file/document
//	                        name (or doc_id when no name is available).
//	file_type     string  — canonical extension used for parser dispatch.
//	output_format string  — always "json".
//	json          []map[string]any — canonical structured parser items.
//	lang          string  — language for tokenization.
//	file          map[string]any — backend-produced file metadata, when present.
//	doc_id        string  — source document ID, when present.
//	bucket        string  — source storage bucket, when present.
//	path          string  — source storage path, when present.
//
// Parser failures are returned as Go errors. The canvas execution wrapper
// preserves that error path and does not convert failures into an _ERROR
// output field.
func (c *ParserComponent) Outputs() map[string]string {
	return map[string]string{
		"name":          "string: the upstream file/document name (or doc_id when no name is available).",
		"file_type":     "string: canonical extension used for parser dispatch (for example pdf, md, xlsx, or other).",
		"output_format": "string: always \"json\".",
		"json":          "[]parser.DocItem: canonical structured parser items.",
		"lang":          "string: the language for tokenization (e.g. English, Dutch, Chinese).",
		"file":          "map[string]any: backend-produced file metadata, when present.",
		"doc_id":        "string: source document ID, when present.",
		"bucket":        "string: source storage bucket, when present.",
		"path":          "string: source storage object path, when present.",
	}
}

// Invoke runs the parser against the upstream "binary" payload.
//
// Returns:
//
//	{
//	  "name":           string (from inputs["name"], file.name, or doc_id),
//	  "file_type":      string (canonical extension used for parser dispatch),
//	  "output_format": "json",
//	  "json":           []parser.DocItem,
//	  "lang":           string (from inputs["lang"]; e.g. English, Dutch),
//	  "_created_time":  RFC3339Nano (via TrackElapsed),
//	  "_elapsed_time":  float64 seconds (via TrackElapsed),
//	}
//
// Per-page parallelism and aggregation now live in the parser
// backends (e.g. internal/deepdoc/parser/pdf fans out one worker
// per page and assembles the results in page order), so this
// component does no goroutine fan-out of its own.
func (c *ParserComponent) Invoke(ctx context.Context, db *gorm.DB, inputs map[string]any) (map[string]any, error) {
	// 1. Decode the binary input.
	binary, err := readParserBinary(ctx, db, inputs)
	if err != nil {
		return nil, err
	}
	docID, _ := inputs["doc_id"].(string)
	filename := parserInputName(inputs, docID)
	setups := c.setups

	// Inject run-level metadata from Globals into inputs so media
	// dispatch branches (audio/image/video) can resolve tenant_id.
	// The File component upstream does not emit tenant_id; the pipeline
	// runner seeds it into CanvasState.Globals, and the Parser must pull
	// it back into the local inputs map for the dispatch functions.
	if tid := globals.GlobalOrInput(ctx, inputs, "tenant_id", ""); tid != "" {
		inputs["tenant_id"] = tid
	}

	// Same pull-back for the run-level (dataset) language: File emits no
	// lang, and the language consumers below — vision enhancement and the
	// media dispatch branches — read the local inputs map, so without this
	// the KB language never reaches them and the prompt language silently
	// falls back to English.
	if lang := globals.GlobalOrInput(ctx, inputs, "lang", ""); lang != "" {
		inputs["lang"] = lang
	}

	// 2. Resolve the file family from the inputs. When the family
	//    is known, dispatchParse returns a typed parser payload.
	//    Otherwise the component stays in text-page mode.
	//
	// We track TWO forms:
	//
	//   - fileTypeExt  — the utility.FileType extension form ("md",
	//     "docx", ...). Used by parser.GetParser, whose switch
	//     arms are keyed off the utility constants.
	//
	requestedType := fileTypeFromInputs(inputs)
	fileParser, fileTypeExt, selectionErr := parser.GetParser(requestedType, binary)
	if selectionErr == nil && fileTypeExt != requestedType {
		warning := fmt.Sprintf("file %q: requested format %s does not match detected format %s; continuing with the %s parser", filename, requestedType, fileTypeExt, fileTypeExt)
		common.Warn(warning)
		reportParserWarnings(ctx, []string{warning})
	}
	parsePagesOn := pdfPageProgressSupported(fileTypeExt, setups)
	enhanceOn := c.enableVisionEnhancement
	split := progress.NewSplit(parsePagesOn, enhanceOn)
	pageWin := split.Take(parsePagesOn)
	visionWin := split.Take(enhanceOn)

	dispatched, handledVision, visionErr := maybeDispatchPDFVision(ctx, db, fileTypeExt, filename, binary, inputs, setups)
	if visionErr != nil {
		return nil, visionErr
	}

	var handledMedia bool
	if !handledVision {
		// Video dispatch: IMAGE2TEXT vision chat.
		// Mirrors Python's _video().
		dispatched, handledMedia, visionErr = maybeDispatchVideo(ctx, db, fileTypeExt, filename, binary, inputs, setups)
		if visionErr != nil {
			return nil, visionErr
		}
	}
	var handledImage bool
	if !handledVision && !handledMedia {
		// Image dispatch: OCR with independently controlled VLM enhancement.
		dispatched, handledImage, visionErr = maybeDispatchImage(ctx, fileParser, db, fileTypeExt, filename, binary, inputs, setups, c.enableVisionEnhancement, c.vision)
		if visionErr != nil {
			return nil, visionErr
		}
	}
	var handledAudio bool
	if !handledVision && !handledMedia && !handledImage {
		// Audio dispatch: SPEECH2TEXT transcription.
		// Mirrors Python's rag/app/audio.py:chunk().
		dispatched, handledAudio, visionErr = maybeDispatchAudio(ctx, db, fileTypeExt, filename, binary, inputs, setups)
		if visionErr != nil {
			return nil, visionErr
		}
	}
	if !handledVision && !handledMedia && !handledImage && !handledAudio {
		if selectionErr != nil && fileTypeExt != utility.FileTypeOTHER {
			return nil, fmt.Errorf("parser: resolve %q: %w", fileTypeExt, selectionErr)
		}
		dispatched = dispatchParse(ctx, fileParser, fileTypeExt, filename, binary, setups, pageWin)
		if dispatched.Err == nil && ctx.Err() == nil {
			pageWin.Report(ctx, 1.0)
		}
	}
	// Known/supported families must fail loudly when dispatch or
	// parsing breaks. Only unknown families keep the raw-text fallback.
	if dispatched.Err != nil && fileTypeExt != utility.FileTypeOTHER {
		return nil, dispatched.Err
	}
	if enhanceOn && !handledMedia && !handledImage && !handledAudio && dispatched.Err == nil && ctx.Err() == nil {
		// PDF dispatch and ordinary parsers share enhancement after normalization.
		var enhanceErr error
		dispatched, _, enhanceErr = maybeDispatchVisionEnhancement(ctx, db, fileTypeExt, dispatched, inputs, c.vision, visionWin)
		if enhanceErr != nil {
			return nil, enhanceErr
		}
	} else {
		// Enhancement skipped entirely (disabled or non-PDF media handled):
		// still strip the transient backend-description marker so it never
		// reaches the chunker.
		dispatched = stripBackendDescribedResult(dispatched)
	}
	reportParserWarnings(ctx, dispatched.Warnings)

	if err := ctx.Err(); err != nil {
		return nil, fmt.Errorf("parser: %w", err)
	}
	lang, _ := getString(inputs, "lang")
	out := buildParserOutputs(ctx, dispatched, filename, binary, lang)

	// Parse billing: DeepDOC PDFs are metered per page after a successful
	// parse. A shortage fails the run, matching the Python hold contract.
	if tenantID, _ := getString(inputs, "tenant_id"); tenantID != "" && docID != "" && fileTypeExt == utility.FileTypePDF {
		if pdfSetup, ok := c.setups["pdf"]; ok {
			if method, _ := pdfSetup["parse_method"].(string); strings.Contains(strings.ToUpper(method), "DEEPDOC") {
				if billingService := billing.Get(); billingService != nil && billingService.Enabled() {
					if runKey := parseRunKey(ctx); runKey != "" {
						if pageCount := parserPageCount(dispatched.File); pageCount > 0 {
							if chargeErr := billingService.ChargeParsedDocument(ctx, tenantID, docID, runKey, pageCount, true); chargeErr != nil {
								return nil, chargeErr
							}
						}
					}
				}
			}
		}
	}
	out["file_type"] = string(fileTypeExt)
	// Forward the storage references so a downstream chunker can
	// re-acquire the source PDF and crop section images on demand,
	// instead of carrying the binary across the component boundary.
	if docID != "" {
		out["doc_id"] = docID
	}
	if bucket, _ := getString(inputs, "bucket"); bucket != "" {
		out["bucket"] = bucket
	}
	if path, _ := getString(inputs, "path"); path != "" {
		out["path"] = path
	}
	// Publish the resolved run-level metadata into the workflow-wide
	// CanvasState.Globals bag so downstream components read it from ctx
	// instead of relying on this output re-emitting it. The Go runtime
	// forwards only this explicit output to the next node, so shared
	// fields must live in Globals.
	globals.PublishGlobals(ctx, out)
	items, _ := out["json"].([]parser.DocItem)
	logParserOutput(dispatched, items)
	// Progress (_created_time / _elapsed_time stamping, start/done
	// callbacks) is owned by the canvas framework (realComponentBody),
	// not by this component, so we return the work result directly.
	return out, nil
}

func logParserOutput(dispatched parser.ParseResult, items []parser.DocItem) {
	common.Debug("parser stage output",
		zap.String("component", "Parser"),
		zap.String("normalized_from", resolveParserNormalizationSource(dispatched)),
		zap.Int("json_items", len(items)),
	)
}

func resolveParserNormalizationSource(dispatched parser.ParseResult) string {
	if len(dispatched.JSON) > 0 {
		return "json"
	}
	return "raw"
}

func reportParserWarnings(ctx context.Context, warnings []string) {
	seen := make(map[string]struct{}, len(warnings))
	for _, warning := range warnings {
		if _, exists := seen[warning]; exists {
			continue
		}
		seen[warning] = struct{}{}
		runtime.ReportProgressMessage(ctx, "Parser", "WARNING: "+warning)
	}
}

// --- input helpers ---

// readParserBinary pulls the "binary" payload out of the inputs
// map. The accepted shapes are:
//
//	[]byte          — the in-process caller's normal form
//	string          — UTF-8 text (JSON callers' normal form)
//	nil / absent    — returns an empty page (not an error)
//
// A non-UTF-8 string is rejected with a clear error so a caller
// that mistakenly hands a base64 string sees the failure
// immediately (mirrors pipeline_chunker's "no try-base64" rule).
func readParserBinary(ctx context.Context, db *gorm.DB, inputs map[string]any) ([]byte, error) {
	if inputs == nil {
		return nil, nil
	}
	if b, ok := inputs["binary"].([]byte); ok {
		return b, nil
	}
	if s, ok := inputs["binary"].(string); ok {
		if !utf8.ValidString(s) {
			return nil, errors.New(
				"parser: binary string is not valid UTF-8. " +
					"Text-page mode only accepts UTF-8 text input")
		}
		return []byte(s), nil
	}
	bucket, _ := getString(inputs, "bucket")
	path, _ := getString(inputs, "path")
	if bucket != "" && path != "" {
		return FetchBinary(ctx, bucket, path)
	}
	if docID, ok := getString(inputs, "doc_id"); ok && docID != "" {
		ref, err := ResolveDocumentStorage(ctx, db, docID)
		if err != nil {
			return nil, fmt.Errorf("parser: resolve doc_id %q: %w", docID, err)
		}
		return FetchBinary(ctx, ref.Bucket, ref.Path)
	}
	return nil, nil
}

// splitIntoPages segments the input bytes on ASCII form-feed
// (\f, 0x0C). An input with no form-feeds becomes a single page
// (the whole input). Empty pages are dropped — the python
// TxtParser skips empty splits the same way.
func splitIntoPages(b []byte) [][]byte {
	if len(b) == 0 {
		return nil
	}
	// Fast path: no form-feeds → single page.
	if !containsFormFeed(b) {
		return [][]byte{b}
	}
	parts := strings.Split(string(b), string(pageFormFeed))
	out := make([][]byte, 0, len(parts))
	for _, p := range parts {
		if len(p) == 0 {
			continue
		}
		out = append(out, []byte(p))
	}
	return out
}

// containsFormFeed is a tiny specialised byte-search to avoid
// pulling in bytes.Index for one call site.
func containsFormFeed(b []byte) bool {
	for _, c := range b {
		if c == pageFormFeed {
			return true
		}
	}
	return false
}

// init registers Parser under CategoryIngestion per plan §4
// Phase 2.2. The factory is a thin closure that decodes the
// DSL param map; the static Metadata is derived from
// Inputs()/Outputs() on a zero-value instance.
func init() {
	pc := &ParserComponent{}
	runtime.MustRegister(ComponentNameParser, runtime.CategoryIngestion,
		func(_ string, params map[string]any) (runtime.Component, error) {
			return NewParserComponent(params)
		},
		runtime.Metadata{
			Version: "1.0.0",
			Inputs:  pc.Inputs(),
			Outputs: pc.Outputs(),
		})
}

// parserPageCount extracts page_count from parser metadata.
func parserPageCount(file map[string]any) int64 {
	if file == nil {
		return 0
	}
	switch value := file["page_count"].(type) {
	case int:
		return int64(value)
	case int64:
		return value
	case float64:
		return int64(value)
	default:
		return 0
	}
}

// parseRunKey returns the billing identity of this parse attempt. The ingestion
// task id is stable across checkpoint resume and component replay, so a retried
// run charges once; only runs without a task scope fall back to a fresh id.
func parseRunKey(ctx context.Context) string {
	if taskID := globals.TaskID(ctx); taskID != "" {
		return "task:" + taskID
	}
	// No task scope means a debug/dry-run preview: nothing is metered.
	return ""
}