Files
cim_summary/backend/src/services/financialTableParser.ts

535 lines
19 KiB
TypeScript

import { logger } from '../utils/logger';
export interface FinancialPeriod {
revenue?: string;
revenueGrowth?: string;
grossProfit?: string;
grossMargin?: string;
ebitda?: string;
ebitdaMargin?: string;
}
export interface ParsedFinancials {
fy3: FinancialPeriod;
fy2: FinancialPeriod;
fy1: FinancialPeriod;
ltm: FinancialPeriod;
}
type Bucket = keyof ParsedFinancials;
const PERIOD_TOKEN_REGEX = /\b(?:(?:FY[-\s]?\d{1,2})|(?:FY[-\s]?)?20\d{2}[A-Z]*|(?:FY[-\s]?[1234])|(?:LTM|TTM))\b/gi;
const MONEY_REGEX = /-?\$?\(?\d[\d,]*(?:\.\d+)?\)?\s?(?:K|M|B)?/g;
const PERCENT_REGEX = /-?\d{1,3}(?:\.\d+)?\s?%/g;
const resetRegex = (regex: RegExp): void => {
regex.lastIndex = 0;
};
const ROW_MATCHERS: Record<string, RegExp> = {
revenue: /(revenue|net sales|total sales|top\s+line)/i,
grossProfit: /(gross\s+profit)/i,
grossMargin: /(gross\s+margin)/i,
ebitda: /(ebitda|adjusted\s+ebitda|adj\.*\s*ebitda)/i,
ebitdaMargin: /(ebitda\s+margin|adj\.*\s*ebitda\s+margin)/i,
revenueGrowth: /(revenue\s+growth|yoy|y\/y|year[-\s]*over[-\s]*year)/i
};
function normalizeToken(token: string): string {
return token.replace(/\s+/g, ' ').replace(/[()]/g, '').trim();
}
function tokenizePeriodHeaders(line: string): string[] {
const matches = line.match(PERIOD_TOKEN_REGEX);
if (!matches) return [];
const normalizedTokens: string[] = [];
for (const match of matches) {
const normalized = normalizePeriodToken(match);
if (!normalized) continue;
if (!normalizedTokens.includes(normalized)) {
normalizedTokens.push(normalized);
}
}
return normalizedTokens;
}
function normalizePeriodToken(rawToken: string): string | null {
if (!rawToken) return null;
const trimmedOriginal = rawToken.trim().toUpperCase();
const isProjection = trimmedOriginal.endsWith('P') || trimmedOriginal.endsWith('PF');
if (isProjection) {
return null;
}
let token = trimmedOriginal.replace(/[\u00A0\s]/g, '');
// Remove trailing punctuation
token = token.replace(/[.,]+$/, '');
// Remove projection suffixes (A, E, F, PF, etc.)
token = token.replace(/(20\d{2})(?:[A-Z]+)$/i, '$1');
token = token.replace(/(FY20\d{2})(?:[A-Z]+)$/i, '$1');
// Normalize FYXX to FY-XX
if (/^FY\d{1,2}$/.test(token)) {
token = token.replace(/^FY(\d{1,2})$/, 'FY-$1');
}
// Normalize FY20XX to just the year
if (/^FY20\d{2}$/.test(token)) {
token = token.replace(/^FY(20\d{2})$/, '$1');
}
return token;
}
function yearTokensToBuckets(tokens: string[]): Array<Bucket | null> {
if (!tokens.length) return [];
const bucketAssignments: Array<Bucket | null> = new Array(tokens.length).fill(null);
const ltmIndices: number[] = [];
// First pass: Identify LTM/TTM periods
tokens.forEach((token, index) => {
if (token.includes('LTM') || token.includes('TTM')) {
bucketAssignments[index] = 'ltm';
ltmIndices.push(index);
}
});
// Get non-LTM indices (these should be fiscal years)
const nonLtmIndices = tokens
.map((token, index) => ({ token, index }))
.filter(({ index }) => !ltmIndices.includes(index));
// Handle edge cases: tables with only 2-3 periods (not all 4)
// Strategy: Assign FY buckets from most recent to oldest (FY1, FY2, FY3)
// If we have 3 years: assign FY1, FY2, FY3
// If we have 2 years: assign FY1, FY2
// If we have 1 year: assign FY1
const fyBuckets: Bucket[] = ['fy1', 'fy2', 'fy3'];
let fyIndex = 0;
// Assign from most recent (rightmost) to oldest (leftmost)
// This matches typical table layout: oldest year on left, newest on right
for (let i = nonLtmIndices.length - 1; i >= 0 && fyIndex < fyBuckets.length; i--) {
const { index } = nonLtmIndices[i];
bucketAssignments[index] = fyBuckets[fyIndex];
fyIndex++;
}
// Validation: Log if we have unusual period counts
const assignedBuckets = bucketAssignments.filter(Boolean);
if (assignedBuckets.length < 2) {
logger.debug('Financial parser: Few periods detected', {
totalTokens: tokens.length,
assignedBuckets: assignedBuckets.length,
tokens: tokens.slice(0, 10)
});
} else if (assignedBuckets.length > 4) {
logger.debug('Financial parser: Many periods detected - may include projections', {
totalTokens: tokens.length,
assignedBuckets: assignedBuckets.length,
tokens: tokens.slice(0, 10)
});
}
return bucketAssignments;
}
/**
* Extract numeric tokens (money/percentages) from a line or combined lines.
* Best practice: Extract all numeric values and preserve their order to match column positions.
*/
function extractNumericTokens(line: string, additionalContent?: string): string[] {
const combined = additionalContent ? `${line} ${additionalContent}` : line;
const lineLength = line.length;
// Extract money values with their positions to preserve column order
resetRegex(MONEY_REGEX);
const moneyMatches = Array.from(combined.matchAll(MONEY_REGEX))
.map((m) => ({ value: normalizeToken(m[0]), index: m.index ?? 0 }))
.filter((m) => m.value && /\d/.test(m.value));
// Extract percentage values with their positions
resetRegex(PERCENT_REGEX);
const percentMatches = Array.from(combined.matchAll(PERCENT_REGEX))
.map((m) => ({ value: normalizeToken(m[0]), index: m.index ?? 0 }))
.filter((m) => m.value && /\d/.test(m.value));
const sortedMatches = [...moneyMatches, ...percentMatches].sort((a, b) => a.index - b.index);
const primaryTokens = sortedMatches
.filter(match => match.index < lineLength)
.map(match => match.value);
if (primaryTokens.length >= 2 || !additionalContent) {
return primaryTokens.length > 0 ? primaryTokens : sortedMatches.map(match => match.value);
}
const secondaryTokens = sortedMatches
.filter(match => match.index >= lineLength)
.map(match => match.value);
return primaryTokens.concat(secondaryTokens);
}
function isMoneyLike(value?: string): boolean {
if (!value) return false;
const clean = value.replace(/[(),\s]/g, '');
return /\d/.test(clean) && (value.includes('$') || /[KMB]/i.test(value));
}
function isPercentLike(value?: string): boolean {
if (!value) return false;
return /\d/.test(value) && value.includes('%');
}
/**
* Assign tokens to buckets based on column position.
* Best practice: Map tokens to buckets by index position, ensuring alignment with header columns.
* This assumes tokens are in the same order as the header columns.
*/
function assignTokensToBuckets(
tokens: string[],
buckets: Array<Bucket | null>,
mapper: (bucket: Bucket, value: string) => void,
fieldName?: string,
lineIndex?: number
) {
// Count non-null buckets (actual periods we want to extract)
const validBuckets = buckets.filter(Boolean).length;
// Validation: Check if token count matches expected bucket count
// Allow some flexibility - tokens can be within 1 of valid buckets (handles missing values)
if (tokens.length < validBuckets - 1) {
logger.debug('Financial parser: Token count mismatch - too few tokens', {
field: fieldName,
lineIndex,
tokensFound: tokens.length,
validBuckets,
tokens: tokens.slice(0, 10),
buckets: buckets.map(b => b || 'skip')
});
// Still try to assign what we have, but log the issue
} else if (tokens.length > validBuckets + 1) {
logger.debug('Financial parser: Token count mismatch - too many tokens', {
field: fieldName,
lineIndex,
tokensFound: tokens.length,
validBuckets,
tokens: tokens.slice(0, 10),
buckets: buckets.map(b => b || 'skip')
});
// Take only the first N tokens that match buckets
}
// Map tokens to buckets by position
// Strategy: Match tokens sequentially to non-null buckets
let tokenIndex = 0;
for (let i = 0; i < buckets.length && tokenIndex < tokens.length; i++) {
const bucket = buckets[i];
if (!bucket) {
// Skip this column (it's a projection or irrelevant period)
// CRITICAL: When we skip a bucket, we also skip the corresponding token
// This assumes tokens are aligned with columns in the table
// If the table has missing values, tokens might be misaligned
// In that case, we try to match by counting non-null buckets before this position
const nonNullBucketsBefore = buckets.slice(0, i).filter(Boolean).length;
if (tokenIndex < nonNullBucketsBefore) {
// We're behind - this might be a missing value, skip the token
tokenIndex++;
}
continue;
}
// Assign the token to this bucket
if (tokenIndex < tokens.length) {
mapper(bucket, tokens[tokenIndex]);
tokenIndex++;
} else {
// No more tokens - this period has no value
logger.debug('Financial parser: Missing token for bucket', {
field: fieldName,
bucket,
bucketIndex: i,
tokensFound: tokens.length
});
}
}
// Log if we didn't use all tokens (might indicate misalignment)
if (tokenIndex < tokens.length && tokens.length > validBuckets) {
logger.debug('Financial parser: Unused tokens detected', {
field: fieldName,
tokensUsed: tokenIndex,
tokensTotal: tokens.length,
validBuckets,
unusedTokens: tokens.slice(tokenIndex)
});
}
}
export function parseFinancialsFromText(fullText: string): ParsedFinancials {
const startTime = Date.now();
const result: ParsedFinancials = {
fy3: {},
fy2: {},
fy1: {},
ltm: {}
};
try {
const text = fullText.replace(/\u00A0/g, ' ');
const lines = text.split('\n').map((line) => line.trim()).filter(Boolean);
if (lines.length === 0) {
return result;
}
let bestHeaderIndex = -1;
let bestBuckets: Array<Bucket | null> = [];
let bestHeaderScore = 0;
// Locate best header line containing year-like tokens
// Best practice: Score headers by both period count AND likelihood of being a financial table
for (let i = 0; i < lines.length; i++) {
const tokens = tokenizePeriodHeaders(lines[i]);
if (tokens.length >= 2) {
const buckets = yearTokensToBuckets(tokens);
const validBuckets = buckets.filter(Boolean).length;
// Score this header: prioritize headers followed by financial metric rows
let score = validBuckets;
// CRITICAL: Financial sections are typically in the BACK HALF of the document
// Boost score for headers in the latter portion of the document
const documentPosition = i / lines.length;
if (documentPosition > 0.5) {
score += 50; // Strong boost for headers in back half
} else if (documentPosition > 0.4) {
score += 20; // Moderate boost for headers in second half
}
// CRITICAL: Financial tables almost always have BOTH revenue AND EBITDA rows
// Look ahead 5-20 lines for these key indicators
const lookAheadStart = Math.min(i + 1, lines.length);
const lookAheadEnd = Math.min(i + 20, lines.length);
let hasRevenue = false;
let hasEBITDA = false;
let financialRowCount = 0;
for (let j = lookAheadStart; j < lookAheadEnd; j++) {
const checkLine = lines[j] || '';
const hasNumbers = containsMoneyOrPercent(checkLine);
if (!hasNumbers) continue; // Skip lines without numbers
// Check for revenue (and variations)
if (ROW_MATCHERS.revenue.test(checkLine)) {
hasRevenue = true;
financialRowCount++;
}
// Check for EBITDA (and variations)
if (ROW_MATCHERS.ebitda.test(checkLine)) {
hasEBITDA = true;
financialRowCount++;
}
// Also count other financial metrics
if (ROW_MATCHERS.grossProfit.test(checkLine) ||
ROW_MATCHERS.grossMargin.test(checkLine) ||
ROW_MATCHERS.ebitdaMargin.test(checkLine) ||
ROW_MATCHERS.revenueGrowth.test(checkLine)) {
financialRowCount++;
}
}
// MASSIVE boost if header has BOTH revenue AND EBITDA (strongest signal)
if (hasRevenue && hasEBITDA) {
score += 100; // This is almost certainly the financial table
} else if (hasRevenue || hasEBITDA) {
score += 20; // Has one key metric
}
// Additional boost for other financial rows
score += financialRowCount * 5;
// Log scoring details for debugging (only for headers with potential)
if (validBuckets >= 2 && (hasRevenue || hasEBITDA || financialRowCount > 0)) {
logger.debug('Financial parser header scoring', {
headerIndex: i,
headerLine: lines[i].substring(0, 100),
validBuckets,
hasRevenue,
hasEBITDA,
financialRowCount,
score,
lookAheadWindow: `${lookAheadStart}-${lookAheadEnd}`
});
}
// Prefer headers with more valid buckets (more historical periods)
if (score > bestHeaderScore || (score === bestHeaderScore && validBuckets > bestBuckets.filter(Boolean).length)) {
bestHeaderScore = score;
bestBuckets = buckets;
bestHeaderIndex = i;
}
}
}
if (bestHeaderIndex === -1 || bestBuckets.filter(Boolean).length === 0) {
logger.info('Financial parser could not identify year header, returning empty result', {
totalLines: lines.length,
sampleLines: lines.slice(0, 20).join(' | ')
});
return result;
}
logger.info('Financial parser selected best header', {
headerIndex: bestHeaderIndex,
headerScore: bestHeaderScore,
buckets: bestBuckets.map((bucket) => bucket || 'skip')
});
logger.info('Financial parser found header', {
headerIndex: bestHeaderIndex,
headerLine: lines[bestHeaderIndex],
buckets: bestBuckets.map((bucket) => bucket || 'skip'),
totalLines: lines.length
});
// Expand window to search for financial data rows (header might be separated from data)
const windowStart = Math.max(0, bestHeaderIndex - 10);
const windowEnd = Math.min(lines.length, bestHeaderIndex + 50); // Increased from 18 to 50 to find data rows
const windowLines = lines.slice(windowStart, windowEnd);
logger.info('Financial parser window', {
windowStart,
windowEnd,
windowSize: windowLines.length,
windowLines: windowLines.join(' | ')
});
const bucketSetters: Record<string, (bucket: Bucket, value: string) => void> = {
revenue: (bucket, value) => {
if (isMoneyLike(value)) result[bucket].revenue = result[bucket].revenue || value;
},
grossProfit: (bucket, value) => {
if (isMoneyLike(value)) result[bucket].grossProfit = result[bucket].grossProfit || value;
},
ebitda: (bucket, value) => {
if (isMoneyLike(value)) result[bucket].ebitda = result[bucket].ebitda || value;
},
grossMargin: (bucket, value) => {
if (isPercentLike(value)) result[bucket].grossMargin = result[bucket].grossMargin || value;
},
ebitdaMargin: (bucket, value) => {
if (isPercentLike(value)) result[bucket].ebitdaMargin = result[bucket].ebitdaMargin || value;
},
revenueGrowth: (bucket, value) => {
if (isPercentLike(value)) result[bucket].revenueGrowth = result[bucket].revenueGrowth || value;
}
};
let matchedRows = 0;
// Search in a larger window around the header for financial data rows
// Also search lines that come after the header (financial tables are usually below headers)
const searchStart = bestHeaderIndex;
const searchEnd = Math.min(lines.length, bestHeaderIndex + 100); // Search up to 100 lines after header
for (let i = searchStart; i < searchEnd; i++) {
const line = lines[i];
if (!line || line.trim().length === 0) continue;
// Check current line and next few lines for numbers (tables might span multiple lines)
const nextLine = lines[i + 1] || '';
const lineAfterNext = lines[i + 2] || '';
const combinedForTokens = `${line} ${nextLine} ${lineAfterNext}`;
// CRITICAL: Only match rows that contain BOTH the field name AND numeric values
// This prevents matching descriptive text that just mentions financial terms
const hasMoneyOrPercent = containsMoneyOrPercent(combinedForTokens);
if (!hasMoneyOrPercent) continue; // Skip lines without actual financial numbers
for (const [field, matcher] of Object.entries(ROW_MATCHERS)) {
if (!matcher.test(line)) continue;
// Extract tokens from the combined lines
const extraContent = `${nextLine} ${lineAfterNext}`.trim() || undefined;
let tokens = extractNumericTokens(line, extraContent);
if (['grossMargin', 'ebitdaMargin', 'revenueGrowth'].includes(field)) {
const percentTokens = tokens.filter(isPercentLike);
if (percentTokens.length > 0) {
tokens = percentTokens;
}
} else if (['revenue', 'grossProfit', 'ebitda'].includes(field)) {
const moneyTokens = tokens.filter(isMoneyLike);
if (moneyTokens.length > 0) {
tokens = moneyTokens;
}
}
// Only process if we found meaningful tokens (at least 2, indicating multiple periods)
if (tokens.length < 2) {
logger.debug('Financial parser: matched field but insufficient tokens', {
field,
lineIndex: i,
tokensFound: tokens.length,
line: line.substring(0, 100)
});
continue;
}
matchedRows++;
logger.info('Financial parser matched row', {
field,
lineIndex: i,
line: line.substring(0, 150),
nextLine: nextLine.substring(0, 100),
tokensFound: tokens.length,
tokens: tokens.slice(0, 10), // Limit token logging
buckets: bestBuckets.map(b => b || 'skip')
});
assignTokensToBuckets(
tokens,
bestBuckets,
(bucket, value) => {
bucketSetters[field](bucket, value);
},
field,
i
);
}
}
logger.info('Financial parser row matching summary', {
matchedRows,
bestBuckets: bestBuckets.length,
buckets: bestBuckets.map((bucket) => bucket || 'skip')
});
logger.info('Financial parser results', {
elapsedMs: Date.now() - startTime,
headerLine: lines[bestHeaderIndex],
fy3: result.fy3,
fy2: result.fy2,
fy1: result.fy1,
ltm: result.ltm
});
} catch (error) {
logger.warn('Financial parser failed', { error: error instanceof Error ? error.message : String(error) });
}
return result;
}
const containsMoneyOrPercent = (text: string): boolean => {
resetRegex(MONEY_REGEX);
const hasMoney = MONEY_REGEX.test(text);
resetRegex(PERCENT_REGEX);
const hasPercent = PERCENT_REGEX.test(text);
return hasMoney || hasPercent;
};