Skip to content

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

18.1 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 chunker

import (
	"crypto/sha256"
	"encoding/hex"
	"encoding/json"
	"fmt"
	"regexp"
	"strings"

	"ragflow/internal/agent/runtime"
	"ragflow/internal/common"
	"ragflow/internal/entity"
	"ragflow/internal/ingestion/component/schema"
	"ragflow/internal/parser/chunk"
	"ragflow/internal/parser/tableutil"
	"ragflow/internal/tokenizer"
)

// newChunkerByName dispatches the DSL name to a typed constructor.
// Centralised here so each chunker file only needs an init() that
// declares its registered name (see register.go). The returned
// runtime.Component interface is satisfied directly by each
// constructor (NewTokenChunker etc.) — no intermediate assertion
// is needed.
func newChunkerByName(name string, params map[string]any) (runtime.Component, error) {
	switch name {
	case ComponentNameTokenChunker:
		// The DSL contract (shared by the web UI and the Python runtime)
		// expresses single-chunk mode as TokenChunker delimiter_mode "one";
		// in Go that behaviour lives in the OneChunker component.
		if mode, _ := params["delimiter_mode"].(string); mode == "one" {
			return NewOneChunker(params)
		}
		return NewTokenChunker(params)
	case ComponentNameTitleChunker:
		return NewTitleChunker(params)
	case ComponentNameGroupTitleChunker:
		return NewGroupTitleChunker(params)
	case ComponentNameManualChunker:
		return NewManualChunker(params)
	case ComponentNameHierarchyTitleChunker:
		return NewHierarchyTitleChunker(params)
	case ComponentNameQAChunker:
		return NewQAChunker(params)
	case ComponentNameOneChunker:
		return NewOneChunker(params)
	case ComponentNameTableChunker:
		return NewTableChunker(params)
	case ComponentNamePageChunker:
		return NewPageChunker(params)
	case ComponentNameGeneralChunker:
		return NewGeneralChunker(params)
	default:
		return nil, fmt.Errorf("chunker: unknown component %q", name)
	}
}

// ---------------------------------------------------------------------------
// numeric / list conversion helpers (shared across chunker variants)
// ---------------------------------------------------------------------------

func stringListFromAny(in []any) []string {
	out := make([]string, 0, len(in))
	for _, x := range in {
		if s, ok := x.(string); ok && s != "" {
			out = append(out, s)
		}
	}
	return out
}

// ---------------------------------------------------------------------------
// regex / split helpers
// ---------------------------------------------------------------------------

// compileDelimPattern compiles a TokenChunker-style []string delimiter list.
// Every non-empty entry is active, including bare (non-backtick) delimiters,
// mirroring Python naive_merge / rag/nlp/delim where bare single-character
// delimiters still split. Backtick-wrapped entries contribute their inner
// content. invokeTextPayload decides whether an active delimiter yields one
// chunk per segment (custom/backtick, no merge) or splits into paragraphs that
// are merged by token size (bare). Canonical single-string parser_config.delimiter
// Legacy single-string parsing is performed by GeneralChunker at its
// configuration boundary; the shared regex helper only consumes canonical
// delimiter lists.
func compileDelimPattern(delims []string) *regexp.Regexp {
	return chunk.CompileDelimiterPatternList(delims, true)
}

// splitByDelim splits text on every match of pattern.
//
// When keepDelim is true the matched delimiter is preserved by gluing it to the
// end of the segment that precedes it, so concatenating the returned segments
// reproduces text exactly: "a。b。" yields ["a。", "b。"] and a leading or
// trailing delimiter stays attached to its neighbour. This is the lossless
// TokenChunker contract — a delimiter is a hint for where the chunker MAY
// break, never a character to delete, so "。" must not vanish (or be swapped
// for "\n") after chunking.
//
// When keepDelim is false the delimiter is discarded. Custom backtick
// delimiters use this mode because they explicitly mark content to split away.
func splitByDelim(text string, pattern *regexp.Regexp, keepDelim bool) []string {
	if pattern == nil {
		return []string{text}
	}
	idxs := pattern.FindAllStringIndex(text, -1)
	if len(idxs) == 0 {
		return []string{text}
	}
	if !keepDelim {
		var out []string
		for _, part := range splitDroppedDelimiterRanges(len(text), idxs) {
			out = append(out, text[part[0]:part[1]])
		}
		return out
	}
	// Keep mode: attach each delimiter to the segment ending just before it so
	// the concatenation is lossless. Trailing text with no trailing delimiter
	// is appended as the final segment.
	var out []string
	cursor := 0
	for _, idx := range idxs {
		start, end := idx[0], idx[1]
		out = append(out, text[cursor:start]+text[start:end])
		cursor = end
	}
	if tail := text[cursor:]; tail != "" {
		out = append(out, tail)
	}
	return out
}

func splitDroppedDelimiterRanges(length int, matches [][]int) [][2]int {
	var ranges [][2]int
	cursor := 0
	for _, match := range matches {
		if match[0] > cursor {
			ranges = append(ranges, [2]int{cursor, match[0]})
		}
		cursor = match[1]
	}
	if cursor < length {
		ranges = append(ranges, [2]int{cursor, length})
	}
	return ranges
}

// foldWhitespaceSegments attaches blank runs to neighboring text without
// changing their position in the reconstructed source.
func foldWhitespaceSegments(parts []string) []string {
	var kept []string
	var leading string
	for _, part := range parts {
		if strings.TrimSpace(part) == "" {
			leading += part
			continue
		}
		if leading != "" {
			part = leading + part
			leading = ""
		}
		kept = append(kept, part)
	}
	if leading != "" && len(kept) > 0 {
		kept[len(kept)-1] += leading
	}
	return kept
}

// ---------------------------------------------------------------------------
// chunk-doc helpers
// ---------------------------------------------------------------------------

// requireChunkText enforces the pre-index wire contract before chunk-id
// generation or image upload.
func requireChunkText(ck map[string]any) (string, error) {
	textRaw, exists := ck["text"]
	if !exists {
		return "", fmt.Errorf("chunk missing required string text field")
	}
	text, ok := textRaw.(string)
	if !ok {
		return "", fmt.Errorf("chunk text must be string, got %T", textRaw)
	}
	return text, nil
}

// itemText returns the canonical pre-index text payload from a chunk item.
func itemText(it schema.ChunkDoc) (string, bool) {
	if it.Text != "" {
		return it.Text, true
	}
	return "", false
}

// itemDocType mirrors _build_json_chunks's type derivation.
func itemDocType(it schema.ChunkDoc) string {
	switch strings.ToLower(strings.TrimSpace(it.DocType)) {
	case "table":
		return "table"
	case "image":
		return "image"
	}
	return "text"
}

// isMediaChunk reports whether a chunk is an image or table region. The
// content type lives on doc_type_kwd; layout never describes media.
func isMediaChunk(ck schema.ChunkDoc) bool {
	return isMediaDocType(ck.DocType)
}

// isMediaDocType reports whether a doc_type_kwd value denotes a media region.
func isMediaDocType(docType string) bool {
	switch strings.ToLower(strings.TrimSpace(docType)) {
	case "image", "table":
		return true
	}
	return false
}

// hasMediaPayload reports whether a chunk still carries retrievable media:
// inline bytes, an uploaded reference, or a deferred media_ref that the
// upload pass will materialize. A present but unusable reference counts:
// dropping the item here would silently lose the image instead of letting
// the materializer reject it.
func hasMediaPayload(ck schema.ChunkDoc) bool {
	return ck.Image != "" || ck.ImgID != "" || ck.MediaRef != nil
}

// itemTextOrFallback returns the item's preferred text, or "".
func itemTextOrFallback(it schema.ChunkDoc) string {
	if t, ok := itemText(it); ok {
		return t
	}
	return ""
}

// materializeTableText fills Text with the HTML rendering of a structured
// TableData when the item carries no text of its own — the shape every parser
// emits for tables. Every consumer then works on a real table it can store and
// display, and the chunk id (a hash of Text) matches the stored content.
// RenderTableHTML marks the leading HeaderRows rows as <th>. Idempotent: an
// item that already carries text is left untouched.
func materializeTableText(docs []schema.ChunkDoc) {
	for i := range docs {
		if docs[i].TableData != nil && docs[i].Text == "" {
			docs[i].Text = tableutil.RenderTableHTML(docs[i].TableData)
		}
	}
}

// tokenizeStr is the shared NumTokensFromString wrapper used by
// Table/Image context attachment. Lives here so we can centrally
// swizzle the count strategy in one place if needed.
func tokenizeStr(s string) int { return tokenizer.NumTokensFromString(s) }

// setChunkText replaces a chunk's body together with the token count that
// describes it, so a caller cannot leave the two out of sync. TKNums is read as
// a budget by the media window walk and by the merge thresholds, and it is
// emitted as tk_nums, so a body replaced on its own silently misreports the
// chunk's size. Merge paths that join several units keep their summed count.
func setChunkText(ck *schema.ChunkDoc, text string) {
	ck.Text = text
	ck.TKNums = intPtr(tokenizeStr(text))
}

// toString normalises a chunk-map field to a string. Empty strings
// for missing fields.
func toString(v any) string {
	if v == nil {
		return ""
	}
	if s, ok := v.(string); ok {
		return s
	}
	return ""
}

// emptyOutputs returns the canonical no-chunks payload.
func emptyOutputs() map[string]any {
	return map[string]any{
		"output_format": "chunks",
		"chunks":        []map[string]any{},
	}
}

// inputErrorOutput is the canonical short-circuit for an undecodable upstream
// payload, shared by every chunker variant's decode step.
func inputErrorOutput(err error) map[string]any {
	return map[string]any{
		"output_format": "chunks",
		"chunks":        []map[string]any{},
		"_ERROR":        fmt.Sprintf("Input error: %v", err),
	}
}

// chunkOutputs builds the canonical chunker output (output_format="chunks" +
// chunks). The Go runtime passes only this explicit output to the next node,
// so the run-level metadata that downstream components still need (e.g.
// `name` for Tokenizer title embedding, or tenant_id/kb_id for embedding
// model resolution) is NOT re-emitted here — it lives in the workflow-wide
// CanvasState.Globals bag (seeded at pipeline start, published by the File
// component) and read directly from ctx. See runtime.CanvasState.Globals.
//
// Media context is materialized into the chunk body here, the chunker's last
// step: every variant passes through this builder, so the folded text is what
// the chunk id, the extractor, the tokenizer and the index write all see.
func chunkOutputs(chunks []schema.ChunkDoc) map[string]any {
	materialized := make([]schema.ChunkDoc, len(chunks))
	for i := range chunks {
		materialized[i] = materializeMediaContext(chunks[i])
		// Structural layout is chunker-internal: it is consumed by merge,
		// split, and crop, and must not cross the output boundary. PDF model
		// layouts (title/figure/equation/...) are not structural and pass
		// through unchanged.
		clearStructuralLayout(&materialized[i])
	}
	return map[string]any{
		"output_format": "chunks",
		"chunks":        schema.ChunkDocsToMaps(materialized),
	}
}

// canonicalChunkText returns the normalized text a chunk id is derived
// from. It folds media context (when present) and strips position tags,
// matching exactly the text the decorator keys on after
// finalizeGeneralChunks runs removeTag and chunkOutputs materializes
// context. Routing every chunk id through this one function — instead of
// recomputing the text at each consumption point — guarantees the streamed
// crop-upload key and the decorator's ck["id"] can never diverge, including
// for image/table chunks whose text still carries position tags when there
// is no media context.
func canonicalChunkText(ck schema.ChunkDoc) string {
	return removeTag(materializeMediaContext(ck).Text)
}

// mediaChunkIDPrefix marks a media-specific identity source. The NUL byte
// keeps the source from colliding with ordinary chunk text.
const mediaChunkIDPrefix = "\x00media:"

// canonicalChunkID derives the crop/upload key from the canonical chunk source.
// The decorator in register.go reuses stamped IDs or derives the same identity
// from finalized text, media payloads and crop regions for chunks without a stamp.
// Title-family media IDs also retain source identity when context or headings
// give distinct captionless media the same text.
func canonicalChunkID(docID string, ck schema.ChunkDoc) string {
	// A title-family captionless image can already carry a source/payload/body
	// identity. Only the title crop bridge marks these finalized IDs; other
	// chunkers may carry upstream IDs whose text they have since rewritten.
	if ck.ID != "" && string(ck.Extra[titleMediaContextIDKey]) == "true" {
		return ck.ID
	}
	return common.ChunkID(docID, chunkIDSource(ck))
}

// chunkIDSource returns the identity string hashed into a chunk id. For a
// media chunk whose canonical text is empty - a captionless image - hashing
// the text alone would give every such chunk in a document the same id
// (ChunkID(docID, "")), silently overwriting images in MinIO and merging
// index entries. Position-only figures use the crop region; otherwise a
// stable per-image payload identity is folded in. Non-empty media text also
// folds its region in: two figures can share a caption (for example "Figure"),
// and their previews must not collapse onto one id. Text chunks hash their text
// unchanged.
func chunkIDSource(ck schema.ChunkDoc) string {
	text := canonicalChunkText(ck)
	matrix := mergePositionMatrix(ck.PDFPositions, ck.Positions)
	if text == "" {
		if isMediaChunk(ck) && len(matrix) > 0 {
			return mediaTextIdentity(text, true, matrix)
		}
		return mediaChunkIdentity(ck.Image, ck.ImgID, ck.MediaRef)
	}
	return mediaTextIdentity(text, isMediaChunk(ck), matrix)
}

// mediaTextIdentity appends a region digest to a media chunk's text identity.
// Non-media chunks and media without a region keep their text-only id.
func mediaTextIdentity(text string, isMedia bool, matrix [][]any) string {
	if !isMedia || len(matrix) == 0 {
		return text
	}
	sum := sha256.Sum256(positionMatrixBytes(matrix))
	return text + mediaChunkIDPrefix + "region:" + hex.EncodeToString(sum[:16])
}

// positionMatrixBytes renders a position matrix deterministically for hashing.
// mergePositionMatrix already sorts the rows, so equal regions always encode
// identically.
func positionMatrixBytes(matrix [][]any) []byte {
	raw, err := json.Marshal(matrix)
	if err != nil {
		return nil
	}
	return raw
}

// chunkPositionMatrix decodes the position rows of a chunk map for identity
// computation, using the same decoder as the crop path.
func chunkPositionMatrix(ck map[string]any) [][]any {
	var sources []json.RawMessage
	for _, key := range []string{"_pdf_positions", "positions"} {
		value, ok := ck[key]
		if !ok || value == nil {
			continue
		}
		if raw, err := json.Marshal(value); err == nil {
			sources = append(sources, raw)
		}
	}
	return mergePositionMatrix(sources...)
}

// mediaChunkIdentity returns the stable identity of a text-less media chunk:
// a digest of the image payload when present (identical images collapse onto
// one id, which is harmless — they carry identical bytes), else the
// content-addressed media reference, else the uploaded image id. Empty when
// the chunk has none of them; callers with PDF crop regions use
// mediaTextIdentity instead.
func mediaChunkIdentity(image, imgID string, ref *entity.MediaRef) string {
	if image != "" {
		sum := sha256.Sum256([]byte(image))
		return mediaChunkIDPrefix + hex.EncodeToString(sum[:])
	}
	if ref.Valid() {
		return mediaChunkIDPrefix + "sha256:" + ref.SHA256
	}
	if imgID != "" {
		return mediaChunkIDPrefix + imgID
	}
	return ""
}

// materializeMediaContext folds a media chunk's surrounding context into its
// body and clears the two fields that carried it. Python's chunker emits the
// same shape — its finalize builds remove_tag(context_above + text +
// context_below) and drops the fields (rag/flow/chunker/token_chunker.py:343-
// 359) — which is why Python persists the context inside the chunk body.
// Folding here also puts the context into the chunk id (ChunkID hashes the
// body), matching Python's id, which hashes the context-bearing body.
//
// Tag stripping runs after the merge, as in Python: the payload was already
// stripped by the chunker, so this only covers the context, which is collected
// from neighbouring units.
//
// The body goes through setChunkText, so TKNums keeps describing the body the
// chunk carries now instead of the bare payload it replaced.
//
// Only media chunks carry context (attachMediaContext and
// attachGeneralMediaContext write it), so text chunks pass through untouched.
func materializeMediaContext(ck schema.ChunkDoc) schema.ChunkDoc {
	if ck.ContextAbove == "" && ck.ContextBelow == "" {
		return ck
	}
	setChunkText(&ck, removeTag(schema.ContextualText(ck)))
	ck.ContextAbove = ""
	ck.ContextBelow = ""
	return ck
}

// withName returns a shallow copy of inputs with name set, so a component can
// guarantee `name` is present on the map it forwards to a decode step without
// mutating the caller's snapshot.
func withName(inputs map[string]any, name string) map[string]any {
	cp := make(map[string]any, len(inputs)+1)
	for k, v := range inputs {
		cp[k] = v
	}
	cp["name"] = name
	return cp
}

// cloneInputs returns a shallow copy of m with room for one extra key. Used to
// inject the Globals-resolved `name` into the decode input without mutating
// the caller's input snapshot.
func cloneInputs(m map[string]any) map[string]any {
	if m == nil {
		return map[string]any{}
	}
	cp := make(map[string]any, len(m)+1)
	for k, v := range m {
		cp[k] = v
	}
	return cp
}