Skip to content

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

9.5 KB
Raw Blame
package reranking

import (
	"context"
	"encoding/json"
	"regexp"
	"strings"

	"github.com/Tencent/WeKnora/internal/logger"
	"github.com/Tencent/WeKnora/internal/types"
)

// --- Passage cleaning for rerank ---
//
// Rerank models work on semantic text similarity. Markdown formatting, raw URLs,
// image references, table separators, and other structural syntax are noise that
// can dilute the semantic signal. The functions below strip this noise before
// passages are sent to the rerank model.

var (
	// reMarkdownImage matches ![alt](url) — the entire construct is noise.
	// URL group supports one level of balanced parentheses.
	reMarkdownImage = regexp.MustCompile(`!\[[^\]]*\]\([^()\s]*(?:\([^)]*\)[^()\s]*)*\)`)
	// reLinkedImage matches [![alt](img_url)](link_url) — unwrap to ![alt](img_url)
	// so that the subsequent reMarkdownImage pass can remove the image.
	reLinkedImage = regexp.MustCompile(
		`\[!\[([^\]]*)\]\(([^()\s]*(?:\([^)]*\)[^()\s]*)*)\)\]` +
			`\([^()\s]*(?:\([^)]*\)[^()\s]*)*\)`,
	)
	// reMarkdownLink matches [text](url) — we keep the text, drop the URL.
	// URL group supports one level of balanced parentheses.
	reMarkdownLink = regexp.MustCompile(`\[([^\]]+)\]\([^()\s]*(?:\([^)]*\)[^()\s]*)*\)`)
	// reRawURL matches standalone http(s) URLs.
	reRawURL = regexp.MustCompile(`https?://[^\s)\]>]+`)
	// reLatexBlock captures the semantic body of block-level LaTeX ($$...$$).
	reLatexBlock = regexp.MustCompile(`(?s)\$\$(.*?)\$\$`)
	// reTableSep matches table separator rows like |---|---|.
	// Uses [ \t] instead of \s to avoid consuming newlines across rows.
	reTableSep = regexp.MustCompile(`(?m)^[ \t]*\|[ \t:|-]+\|[ \t]*$`)
	// reTableRow matches markdown table data rows like | col1 | col2 |.
	// Uses [ \t] instead of \s to avoid consuming newlines across rows.
	reTableRow = regexp.MustCompile(`(?m)^[ \t]*\|(.+?)\|[ \t]*$`)
	// reHeadingPrefix matches leading # markers in headings.
	reHeadingPrefix = regexp.MustCompile(`(?m)^#{1,6}\s+`)
	// reBlockquote matches leading > markers.
	reBlockquote = regexp.MustCompile(`(?m)^>\s?`)
	// reBoldItalic3 matches ***text*** wrappers (must come before 2 and 1).
	reBoldItalic3 = regexp.MustCompile(`\*{3}(.+?)\*{3}`)
	// reBoldItalic2 matches **text** wrappers.
	reBoldItalic2 = regexp.MustCompile(`\*{2}(.+?)\*{2}`)
	// reBoldItalic1 matches *text* wrappers.
	reBoldItalic1 = regexp.MustCompile(`\*(.+?)\*`)
	// reExcessiveNewlines collapses 3+ consecutive newlines into 2.
	reExcessiveNewlines = regexp.MustCompile(`\n{3,}`)
	// reListMarker matches unordered (- , * ) and ordered (1. ) list prefixes.
	reListMarker = regexp.MustCompile(`(?m)^[\t ]*(?:[-*+]|\d+\.)\s+`)
	// reHTMLTag matches HTML tags like <br>, <div class="...">, etc.
	reHTMLTag = regexp.MustCompile(`</?[a-zA-Z][^>]*>`)
)

// CleanPassage strips markdown/structural noise from text to produce a clean
// semantic passage for the rerank model. The cleaning is designed to preserve
// all meaningful semantic content while removing formatting that would
// confuse text-similarity scoring.
func CleanPassage(text string) string {
	var out, prose strings.Builder
	var fenceChar byte
	fenceLength := 0
	flushProse := func() {
		if prose.Len() == 0 {
			return
		}
		out.WriteString(cleanPassageProse(prose.String()))
		prose.Reset()
	}

	// Keep fenced code separate from prose throughout cleaning. Unwrapping
	// first would let later passes delete XML, URLs and operators in code.
	for line := range strings.SplitAfterSeq(text, "\n") {
		marker, length, tail := passageFence(line)
		if fenceLength == 0 {
			if length < 3 || (marker == '`' && strings.ContainsRune(tail, '`')) {
				prose.WriteString(line)
				continue
			}
			flushProse()
			fenceChar, fenceLength = marker, length
		} else if marker == fenceChar && length >= fenceLength && strings.Trim(tail, " \t\r\n") == "" {
			fenceLength = 0
		} else {
			out.WriteString(line)
		}
	}
	flushProse()
	return strings.TrimSpace(out.String())
}

// passageFence recognizes a top-level fence with up to three leading spaces.
// Matching type/length and a whitespace-only closing tail follow CommonMark
// 4.5. This helper does not recognize container, indented or inline code.
func passageFence(line string) (marker byte, length int, tail string) {
	trimmed := strings.TrimLeft(line, " ")
	if len(line)-len(trimmed) > 3 || len(trimmed) == 0 {
		return 0, 0, ""
	}
	marker = trimmed[0]
	if marker != '`' && marker != '~' {
		return 0, 0, ""
	}
	for length < len(trimmed) && trimmed[length] == marker {
		length++
	}
	return marker, length, trimmed[length:]
}

// cleanPassageProse leaves boundary whitespace intact so code and surrounding
// text retain their line boundaries when the cleaned segments are assembled.
func cleanPassageProse(text string) string {
	// 2. Unwrap LaTeX blocks so formula-only candidates remain rerankable.
	text = reLatexBlock.ReplaceAllString(text, "$1")
	// 3. Remove HTML tags
	text = reHTMLTag.ReplaceAllString(text, "")
	// 3.5. Unwrap nested [![alt](img_url)](link_url) → ![alt](img_url)
	//      so that the next step removes the full construct cleanly.
	text = reLinkedImage.ReplaceAllString(text, "![$1]($2)")
	// 4. Remove markdown image references entirely
	text = reMarkdownImage.ReplaceAllString(text, "")
	// 5. Convert markdown links to just their display text
	text = reMarkdownLink.ReplaceAllString(text, "$1")
	// 6. Remove standalone raw URLs
	text = reRawURL.ReplaceAllString(text, "")
	// 7. Remove table separator rows
	text = reTableSep.ReplaceAllString(text, "")
	// 7.5. Convert table data rows to plain text (strip | delimiters)
	text = reTableRow.ReplaceAllStringFunc(text, func(match string) string {
		inner := reTableRow.FindStringSubmatch(match)
		if len(inner) < 2 {
			return match
		}
		cells := strings.Split(inner[1], "|")
		var parts []string
		for _, cell := range cells {
			cell = strings.TrimSpace(cell)
			if cell != "" {
				parts = append(parts, cell)
			}
		}
		return strings.Join(parts, ", ")
	})
	// 8. Strip heading markers but keep heading text
	text = reHeadingPrefix.ReplaceAllString(text, "")
	// 9. Strip blockquote markers
	text = reBlockquote.ReplaceAllString(text, "")
	// 10. Unwrap bold/italic markers, keeping inner text (order: *** before ** before *)
	text = reBoldItalic3.ReplaceAllString(text, "$1")
	text = reBoldItalic2.ReplaceAllString(text, "$1")
	text = reBoldItalic1.ReplaceAllString(text, "$1")
	// 11. Strip list markers
	text = reListMarker.ReplaceAllString(text, "")
	// 12. Collapse excessive newlines
	text = reExcessiveNewlines.ReplaceAllString(text, "\n\n")

	return text
}

// EnrichedPassage is the cleaned chunk body plus the image captions, OCR text
// and generated questions attached to it. MMR compares candidates on this
// text, so chunks of one document are not made similar by a shared title.
// Generated questions are omitted when the knowledge base has turned question
// generation off.
func EnrichedPassage(ctx context.Context, result *types.SearchResult) string {
	combinedText := CleanPassage(result.Content)
	var enrichments []string
	// An image OCR or caption chunk's body is that same text, and its
	// image_info repeats it; appending it again doubled the passage.
	body := strings.TrimSpace(result.Content)
	addImageText := func(text string) {
		if text != "" && strings.TrimSpace(text) != body {
			enrichments = append(enrichments, text)
		}
	}

	if result.ImageInfo != "" {
		var imageInfos []types.ImageInfo
		if err := json.Unmarshal([]byte(result.ImageInfo), &imageInfos); err != nil {
			logger.Warnf(ctx, "[Rerank] Failed to parse image info of chunk %s: %v", result.ID, err)
		} else {
			for _, img := range imageInfos {
				addImageText(img.Caption)
				addImageText(img.OCRText)
			}
		}
	}

	if len(result.ChunkMetadata) > 0 {
		var docMeta types.DocumentChunkMetadata
		if err := json.Unmarshal(result.ChunkMetadata, &docMeta); err != nil {
			logger.Warnf(ctx, "[Rerank] Failed to parse chunk metadata of chunk %s: %v", result.ID, err)
		} else if includeGeneratedQuestions(result) {
			if questionStrings := docMeta.GetQuestionStrings(); len(questionStrings) > 0 {
				enrichments = append(enrichments, strings.Join(questionStrings, "; "))
			}
		}
	}

	if len(enrichments) == 0 {
		return combinedText
	}
	if combinedText != "" {
		combinedText += "\n\n"
	}
	return combinedText + strings.Join(enrichments, "\n")
}

func includeGeneratedQuestions(result *types.SearchResult) bool {
	return result == nil || !result.SkipGeneratedQuestions
}

// ModelPassage is the text the rerank model scores: the document title
// followed by the enriched chunk. A chunk rarely restates what its document
// is about, so without the title a passage from "Show HN: Echo - ... using
// open-weight models" scored 0.002 against "Echo open-weight models reduce
// cost" and 0.45 with it. FAQ entries carry their own question instead.
//
// An empty enriched body stays empty: a title alone is not evidence that
// the chunk answers anything.
//
// The chunk's heading breadcrumb (ContextHeader) goes between the two for the
// same reason: it was embedded with the chunk, so vector search can find
// "7 天内可申请" under "## 退款政策" while a model scoring the bare body
// rejects it.
func ModelPassage(ctx context.Context, result *types.SearchResult) string {
	passage := EnrichedPassage(ctx, result)
	if strings.TrimSpace(passage) == "" || result.ChunkType == string(types.ChunkTypeFAQ) {
		return passage
	}
	if header := strings.TrimSpace(result.ContextHeader); header != "" {
		passage = header + "\n\n" + passage
	}
	if title := strings.TrimSpace(result.KnowledgeTitle); title != "" {
		passage = title + "\n\n" + passage
	}
	return passage
}