@retrivora-ai/rag-engine 2.5.8 → 2.6.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/dist/server.mjs CHANGED
@@ -326,7 +326,7 @@ var init_package = __esm({
326
326
  "package.json"() {
327
327
  package_default = {
328
328
  name: "@retrivora-ai/rag-engine",
329
- version: "2.5.8",
329
+ version: "2.6.0",
330
330
  description: "Retrivora AI is a plug-and-play AI engine for RAG chat experiences \u2014 generic vector DB + LLM provider, embeddable or standalone.",
331
331
  author: "Abhinav Alkuchi",
332
332
  license: "UNLICENSED",
@@ -2795,7 +2795,7 @@ var _LicenseVerifier = class _LicenseVerifier {
2795
2795
  */
2796
2796
  static async generateMachineFingerprint() {
2797
2797
  if (typeof window !== "undefined" && typeof navigator !== "undefined") {
2798
- return navigator.userAgent || "browser-fallback-fingerprint";
2798
+ return "browser-client";
2799
2799
  }
2800
2800
  try {
2801
2801
  const os2 = await import("os");
@@ -2867,7 +2867,7 @@ var _LicenseVerifier = class _LicenseVerifier {
2867
2867
  try {
2868
2868
  response = await fetch(endpoint, {
2869
2869
  method: "POST",
2870
- signal: AbortSignal.timeout(3e3),
2870
+ signal: AbortSignal.timeout(6e3),
2871
2871
  headers: {
2872
2872
  "Content-Type": "application/json",
2873
2873
  "x-license-key": licenseKey
@@ -5645,27 +5645,384 @@ Text to extract from:
5645
5645
  }
5646
5646
  };
5647
5647
 
5648
+ // src/rag/HybridSearch.ts
5649
+ var HybridSearch = class {
5650
+ /**
5651
+ * Tokenize text into normalized alphanumeric keywords, keeping exact codes, numbers, and identifiers.
5652
+ */
5653
+ static tokenize(text) {
5654
+ if (!text) return [];
5655
+ return text.toLowerCase().replace(/[^\w\s\-_#./:]/g, " ").split(/\s+/).map((t) => t.trim()).filter((t) => t.length > 1 && !this.stopWords.has(t));
5656
+ }
5657
+ /**
5658
+ * Calculate Okapi BM25 sparse scores for an array of document chunks given a query.
5659
+ */
5660
+ static scoreBM25(documents, query, k1 = 1.5, b = 0.75) {
5661
+ const scores = /* @__PURE__ */ new Map();
5662
+ const queryTokens = this.tokenize(query);
5663
+ if (queryTokens.length === 0 || documents.length === 0) {
5664
+ documents.forEach((d) => scores.set(d.id, 0));
5665
+ return scores;
5666
+ }
5667
+ const docTokens = documents.map((d) => ({
5668
+ id: d.id,
5669
+ tokens: this.tokenize(d.content)
5670
+ }));
5671
+ const N = documents.length;
5672
+ const avgdl = docTokens.reduce((acc, d) => acc + d.tokens.length, 0) / (N || 1);
5673
+ const dfMap = /* @__PURE__ */ new Map();
5674
+ for (const qt of queryTokens) {
5675
+ let count = 0;
5676
+ for (const dt of docTokens) {
5677
+ if (dt.tokens.includes(qt)) count++;
5678
+ }
5679
+ dfMap.set(qt, count);
5680
+ }
5681
+ for (const dt of docTokens) {
5682
+ let score = 0;
5683
+ const docLen = dt.tokens.length;
5684
+ const tfMap = /* @__PURE__ */ new Map();
5685
+ for (const t of dt.tokens) {
5686
+ tfMap.set(t, (tfMap.get(t) || 0) + 1);
5687
+ }
5688
+ for (const qt of queryTokens) {
5689
+ const tf = tfMap.get(qt) || 0;
5690
+ if (tf === 0) continue;
5691
+ const df = dfMap.get(qt) || 0;
5692
+ const idf = Math.log((N - df + 0.5) / (df + 0.5) + 1);
5693
+ const numerator = tf * (k1 + 1);
5694
+ const denominator = tf + k1 * (1 - b + b * (docLen / (avgdl || 1)));
5695
+ score += idf * (numerator / (denominator || 1));
5696
+ }
5697
+ scores.set(dt.id, score);
5698
+ }
5699
+ return scores;
5700
+ }
5701
+ /**
5702
+ * Fuse Dense Vector Matches with Sparse BM25 Scores using Reciprocal Rank Fusion (RRF).
5703
+ */
5704
+ static fuse(denseMatches, query, options = {}) {
5705
+ var _a2, _b, _c;
5706
+ if (denseMatches.length === 0) return [];
5707
+ const denseWeight = (_a2 = options.denseWeight) != null ? _a2 : 0.6;
5708
+ const sparseWeight = (_b = options.sparseWeight) != null ? _b : 0.4;
5709
+ const k = (_c = options.rrfK) != null ? _c : 60;
5710
+ const sortedDense = [...denseMatches].sort((a, b) => b.score - a.score);
5711
+ const denseRankMap = /* @__PURE__ */ new Map();
5712
+ sortedDense.forEach((match, idx) => {
5713
+ denseRankMap.set(match.id, idx + 1);
5714
+ });
5715
+ const docItems = denseMatches.map((m) => ({ id: m.id, content: m.content }));
5716
+ const bm25ScoreMap = this.scoreBM25(docItems, query, options.bm25K1, options.bm25B);
5717
+ const sortedSparse = [...denseMatches].sort((a, b) => {
5718
+ const scoreA = bm25ScoreMap.get(a.id) || 0;
5719
+ const scoreB = bm25ScoreMap.get(b.id) || 0;
5720
+ return scoreB - scoreA;
5721
+ });
5722
+ const sparseRankMap = /* @__PURE__ */ new Map();
5723
+ sortedSparse.forEach((match, idx) => {
5724
+ sparseRankMap.set(match.id, idx + 1);
5725
+ });
5726
+ const fusedMatches = denseMatches.map((match) => {
5727
+ var _a3;
5728
+ const denseRank = denseRankMap.get(match.id) || denseMatches.length;
5729
+ const sparseRank = sparseRankMap.get(match.id) || denseMatches.length;
5730
+ const denseRrf = denseWeight / (k + denseRank);
5731
+ const sparseRrf = sparseWeight / (k + sparseRank);
5732
+ const combinedRrf = denseRrf + sparseRrf;
5733
+ const bm25Raw = bm25ScoreMap.get(match.id) || 0;
5734
+ return __spreadProps(__spreadValues({}, match), {
5735
+ rrfScore: combinedRrf,
5736
+ bm25Score: bm25Raw,
5737
+ rerankScore: ((_a3 = match.rerankScore) != null ? _a3 : match.score) * 0.5 + combinedRrf * 50
5738
+ });
5739
+ });
5740
+ return fusedMatches.sort((a, b) => {
5741
+ var _a3, _b2;
5742
+ return ((_a3 = b.rerankScore) != null ? _a3 : b.score) - ((_b2 = a.rerankScore) != null ? _b2 : a.score);
5743
+ });
5744
+ }
5745
+ };
5746
+ HybridSearch.stopWords = /* @__PURE__ */ new Set([
5747
+ "a",
5748
+ "about",
5749
+ "above",
5750
+ "after",
5751
+ "again",
5752
+ "against",
5753
+ "all",
5754
+ "am",
5755
+ "an",
5756
+ "and",
5757
+ "any",
5758
+ "are",
5759
+ "aren't",
5760
+ "as",
5761
+ "at",
5762
+ "be",
5763
+ "because",
5764
+ "been",
5765
+ "before",
5766
+ "being",
5767
+ "below",
5768
+ "between",
5769
+ "both",
5770
+ "but",
5771
+ "by",
5772
+ "can't",
5773
+ "cannot",
5774
+ "could",
5775
+ "couldn't",
5776
+ "did",
5777
+ "didn't",
5778
+ "do",
5779
+ "does",
5780
+ "doesn't",
5781
+ "doing",
5782
+ "don't",
5783
+ "down",
5784
+ "during",
5785
+ "each",
5786
+ "few",
5787
+ "for",
5788
+ "from",
5789
+ "further",
5790
+ "had",
5791
+ "hadn't",
5792
+ "has",
5793
+ "hasn't",
5794
+ "have",
5795
+ "haven't",
5796
+ "having",
5797
+ "he",
5798
+ "he'd",
5799
+ "he'll",
5800
+ "he's",
5801
+ "her",
5802
+ "here",
5803
+ "here's",
5804
+ "hers",
5805
+ "herself",
5806
+ "him",
5807
+ "himself",
5808
+ "his",
5809
+ "how",
5810
+ "how's",
5811
+ "i",
5812
+ "i'd",
5813
+ "i'll",
5814
+ "i'm",
5815
+ "i've",
5816
+ "if",
5817
+ "in",
5818
+ "into",
5819
+ "is",
5820
+ "isn't",
5821
+ "it",
5822
+ "it's",
5823
+ "its",
5824
+ "itself",
5825
+ "let's",
5826
+ "me",
5827
+ "more",
5828
+ "most",
5829
+ "mustn't",
5830
+ "my",
5831
+ "myself",
5832
+ "no",
5833
+ "nor",
5834
+ "not",
5835
+ "of",
5836
+ "off",
5837
+ "on",
5838
+ "once",
5839
+ "only",
5840
+ "or",
5841
+ "other",
5842
+ "ought",
5843
+ "our",
5844
+ "ours",
5845
+ "ourselves",
5846
+ "out",
5847
+ "over",
5848
+ "own",
5849
+ "same",
5850
+ "shan't",
5851
+ "she",
5852
+ "she'd",
5853
+ "she'll",
5854
+ "she's",
5855
+ "should",
5856
+ "shouldn't",
5857
+ "so",
5858
+ "some",
5859
+ "such",
5860
+ "than",
5861
+ "that",
5862
+ "that's",
5863
+ "the",
5864
+ "their",
5865
+ "theirs",
5866
+ "them",
5867
+ "themselves",
5868
+ "then",
5869
+ "there",
5870
+ "there's",
5871
+ "these",
5872
+ "they",
5873
+ "they'd",
5874
+ "they'll",
5875
+ "they're",
5876
+ "they've",
5877
+ "this",
5878
+ "those",
5879
+ "through",
5880
+ "to",
5881
+ "too",
5882
+ "under",
5883
+ "until",
5884
+ "up",
5885
+ "very",
5886
+ "was",
5887
+ "wasn't",
5888
+ "we",
5889
+ "we'd",
5890
+ "we'll",
5891
+ "we're",
5892
+ "we've",
5893
+ "were",
5894
+ "weren't",
5895
+ "what",
5896
+ "what's",
5897
+ "when",
5898
+ "when's",
5899
+ "where",
5900
+ "where's",
5901
+ "which",
5902
+ "while",
5903
+ "who",
5904
+ "who's",
5905
+ "whom",
5906
+ "why",
5907
+ "why's",
5908
+ "with",
5909
+ "won't",
5910
+ "would",
5911
+ "wouldn't",
5912
+ "you",
5913
+ "you'd",
5914
+ "you'll",
5915
+ "you're",
5916
+ "you've",
5917
+ "your",
5918
+ "yours",
5919
+ "yourself",
5920
+ "yourselves"
5921
+ ]);
5922
+
5648
5923
  // src/rag/Reranker.ts
5649
5924
  var MAX_RERANK_CONTENT_CHARS = 64e3;
5650
5925
  var Reranker = class {
5651
5926
  /**
5652
- * Re-ranks matches based on a secondary relevance score using an LLM.
5653
- *
5654
- * Fix RAG-2: keyword re-ranking no longer mutates the original `score`
5655
- * field (which consumers interpret as the raw vector similarity). A new
5656
- * `rerankScore` field is added so both the original similarity and the
5657
- * boosted rank signal are available to callers.
5927
+ * Re-ranks matches based on a secondary relevance score using Hybrid BM25, Cohere, Jina or LLM.
5658
5928
  */
5659
- async rerank(matches, query, limit = 5, llm) {
5660
- if (!llm || matches.length <= 1) {
5661
- return this.keywordRerank(matches, query, limit);
5929
+ async rerank(matches, query, limit = 5, llm, options = {}) {
5930
+ var _a2, _b;
5931
+ if (!matches || matches.length <= 1) {
5932
+ return matches ? matches.slice(0, limit) : [];
5662
5933
  }
5663
- try {
5664
- return await this.llmRerank(matches, query, limit, llm);
5665
- } catch (error) {
5666
- console.warn("[Reranker] LLM re-ranking failed, falling back to keyword sorting:", error);
5667
- return this.keywordRerank(matches, query, limit);
5934
+ const cohereKey = options.cohereApiKey || (typeof process !== "undefined" ? (_a2 = process.env) == null ? void 0 : _a2.COHERE_API_KEY : void 0);
5935
+ const jinaKey = options.jinaApiKey || (typeof process !== "undefined" ? (_b = process.env) == null ? void 0 : _b.JINA_API_KEY : void 0);
5936
+ if ((options.provider === "cohere" || cohereKey) && cohereKey) {
5937
+ try {
5938
+ const cohereResult = await this.cohereRerank(matches, query, limit, cohereKey, options.cohereModel);
5939
+ if (cohereResult.length > 0) return cohereResult;
5940
+ } catch (err) {
5941
+ console.warn("[Reranker] Cohere reranking failed, falling back to HybridSearch:", err);
5942
+ }
5668
5943
  }
5944
+ if ((options.provider === "jina" || jinaKey) && jinaKey) {
5945
+ try {
5946
+ const jinaResult = await this.jinaRerank(matches, query, limit, jinaKey, options.jinaModel);
5947
+ if (jinaResult.length > 0) return jinaResult;
5948
+ } catch (err) {
5949
+ console.warn("[Reranker] Jina reranking failed, falling back to HybridSearch:", err);
5950
+ }
5951
+ }
5952
+ if (!options.provider || options.provider === "hybrid") {
5953
+ const fused = HybridSearch.fuse(matches, query);
5954
+ if (fused.length > 0) {
5955
+ return fused.slice(0, limit);
5956
+ }
5957
+ }
5958
+ if (llm && (options.provider === "llm" || !options.provider)) {
5959
+ try {
5960
+ return await this.llmRerank(matches, query, limit, llm);
5961
+ } catch (error) {
5962
+ console.warn("[Reranker] LLM re-ranking failed, falling back to keyword sorting:", error);
5963
+ }
5964
+ }
5965
+ return this.keywordRerank(matches, query, limit);
5966
+ }
5967
+ /**
5968
+ * Cohere Rerank v3.5 API integration.
5969
+ */
5970
+ async cohereRerank(matches, query, limit, apiKey, model = "rerank-v3.5") {
5971
+ const documents = matches.map((m) => m.content || "");
5972
+ const response = await fetch("https://api.cohere.com/v1/rerank", {
5973
+ method: "POST",
5974
+ headers: {
5975
+ "Authorization": `Bearer ${apiKey}`,
5976
+ "Content-Type": "application/json"
5977
+ },
5978
+ body: JSON.stringify({
5979
+ model,
5980
+ query,
5981
+ documents,
5982
+ top_n: limit
5983
+ })
5984
+ });
5985
+ if (!response.ok) {
5986
+ throw new Error(`Cohere API returned status ${response.status}`);
5987
+ }
5988
+ const data = await response.json();
5989
+ if (!data.results || !Array.isArray(data.results)) return [];
5990
+ return data.results.map((res) => {
5991
+ const match = matches[res.index];
5992
+ return __spreadProps(__spreadValues({}, match), {
5993
+ rerankScore: res.relevance_score
5994
+ });
5995
+ });
5996
+ }
5997
+ /**
5998
+ * Jina AI Reranker v2 API integration.
5999
+ */
6000
+ async jinaRerank(matches, query, limit, apiKey, model = "jina-reranker-v2-base-multilingual") {
6001
+ const documents = matches.map((m) => m.content || "");
6002
+ const response = await fetch("https://api.jina.ai/v1/rerank", {
6003
+ method: "POST",
6004
+ headers: {
6005
+ "Authorization": `Bearer ${apiKey}`,
6006
+ "Content-Type": "application/json"
6007
+ },
6008
+ body: JSON.stringify({
6009
+ model,
6010
+ query,
6011
+ documents,
6012
+ top_n: limit
6013
+ })
6014
+ });
6015
+ if (!response.ok) {
6016
+ throw new Error(`Jina API returned status ${response.status}`);
6017
+ }
6018
+ const data = await response.json();
6019
+ if (!data.results || !Array.isArray(data.results)) return [];
6020
+ return data.results.map((res) => {
6021
+ const match = matches[res.index];
6022
+ return __spreadProps(__spreadValues({}, match), {
6023
+ rerankScore: res.relevance_score
6024
+ });
6025
+ });
5669
6026
  }
5670
6027
  /**
5671
6028
  * Keyword-based re-ranking fallback.
@@ -5751,6 +6108,176 @@ ${docList}` }],
5751
6108
  }
5752
6109
  };
5753
6110
 
6111
+ // src/rag/RAGEvaluator.ts
6112
+ var RAGEvaluator = class {
6113
+ /**
6114
+ * Evaluate a RAG interaction across the full RAG Triad.
6115
+ */
6116
+ static async evaluate(query, context, answer, llm) {
6117
+ if (!llm || !context || !answer) {
6118
+ return this.deterministicHeuristic(query, context, answer);
6119
+ }
6120
+ try {
6121
+ const prompt = `You are an automated RAG Triad Evaluator. Evaluate the quality of the following RAG system interaction:
6122
+
6123
+ USER QUERY:
6124
+ ${query}
6125
+
6126
+ RETRIEVED CONTEXT:
6127
+ ${context.slice(0, 3e3)}
6128
+
6129
+ GENERATED ANSWER:
6130
+ ${answer}
6131
+
6132
+ Evaluate these 3 metrics on a scale of 0.0 to 1.0:
6133
+ 1. "contextRelevance": Are the retrieved documents relevant to the query? (1.0 = highly relevant, 0.0 = irrelevant)
6134
+ 2. "faithfulness": Is every claim in the generated answer supported by the retrieved context? (1.0 = 100% grounded/no hallucination, 0.0 = completely fabricated)
6135
+ 3. "answerRelevance": Does the answer directly address the user's question? (1.0 = direct answer, 0.0 = evasive/unrelated)
6136
+
6137
+ Return ONLY valid JSON with no markdown fences:
6138
+ {"contextRelevance": 0.95, "faithfulness": 0.90, "answerRelevance": 0.92, "reason": "Answer is fully supported by the document."}`;
6139
+ const raw = await llm.chat(
6140
+ [{ role: "user", content: prompt }],
6141
+ "",
6142
+ {
6143
+ systemPrompt: "You are a strict AI safety and hallucination evaluator. Output only valid JSON.",
6144
+ temperature: 0,
6145
+ maxTokens: 150
6146
+ }
6147
+ );
6148
+ const cleaned = raw.replace(/```json|```/g, "").trim();
6149
+ const parsed = JSON.parse(cleaned);
6150
+ const cr = typeof parsed.contextRelevance === "number" ? Math.max(0, Math.min(1, parsed.contextRelevance)) : 0.85;
6151
+ const f = typeof parsed.faithfulness === "number" ? Math.max(0, Math.min(1, parsed.faithfulness)) : 0.9;
6152
+ const ar = typeof parsed.answerRelevance === "number" ? Math.max(0, Math.min(1, parsed.answerRelevance)) : 0.88;
6153
+ const overall = Number((cr * 0.3 + f * 0.4 + ar * 0.3).toFixed(2));
6154
+ return {
6155
+ contextRelevance: cr,
6156
+ faithfulness: f,
6157
+ answerRelevance: ar,
6158
+ overallQuality: overall,
6159
+ isHallucinated: f < 0.6,
6160
+ reason: parsed.reason || "Evaluated successfully."
6161
+ };
6162
+ } catch (err) {
6163
+ return this.deterministicHeuristic(query, context, answer);
6164
+ }
6165
+ }
6166
+ /**
6167
+ * Deterministic fast keyword-overlap heuristic when LLM evaluation is not invoked.
6168
+ */
6169
+ static deterministicHeuristic(query, context, answer) {
6170
+ if (!context || !answer) {
6171
+ return { contextRelevance: 0.5, faithfulness: 0.8, answerRelevance: 0.7, overallQuality: 0.67, isHallucinated: false };
6172
+ }
6173
+ const queryWords = query.toLowerCase().split(/\s+/).filter((w) => w.length > 2);
6174
+ const contextLower = context.toLowerCase();
6175
+ const answerLower = answer.toLowerCase();
6176
+ let queryContextHits = 0;
6177
+ queryWords.forEach((w) => {
6178
+ if (contextLower.includes(w)) queryContextHits++;
6179
+ });
6180
+ const contextRelevance = queryWords.length > 0 ? Math.min(1, queryContextHits / queryWords.length + 0.3) : 0.8;
6181
+ const numbersInAnswer = answer.match(/\b\d+(?:\.\d+)?%?\b/g) || [];
6182
+ let numberHits = 0;
6183
+ numbersInAnswer.forEach((num) => {
6184
+ if (context.includes(num)) numberHits++;
6185
+ });
6186
+ const faithfulness = numbersInAnswer.length > 0 ? Math.min(1, numberHits / numbersInAnswer.length + 0.2) : 0.9;
6187
+ const answerRelevance = 0.85;
6188
+ const overallQuality = Number((contextRelevance * 0.3 + faithfulness * 0.4 + answerRelevance * 0.3).toFixed(2));
6189
+ return {
6190
+ contextRelevance: Number(contextRelevance.toFixed(2)),
6191
+ faithfulness: Number(faithfulness.toFixed(2)),
6192
+ answerRelevance,
6193
+ overallQuality,
6194
+ isHallucinated: faithfulness < 0.5,
6195
+ reason: "Computed via deterministic grounding heuristic."
6196
+ };
6197
+ }
6198
+ };
6199
+
6200
+ // src/security/PIIMasker.ts
6201
+ var PIIMasker = class {
6202
+ /**
6203
+ * Redact sensitive information from text.
6204
+ */
6205
+ static mask(text, options = {}) {
6206
+ if (!text || typeof text !== "string") return text;
6207
+ let sanitized = text;
6208
+ if (options.maskApiKeys !== false) {
6209
+ sanitized = sanitized.replace(this.API_KEY_REGEX, "[REDACTED_API_KEY]");
6210
+ }
6211
+ if (options.maskCreditCards !== false) {
6212
+ sanitized = sanitized.replace(this.CREDIT_CARD_REGEX, (match) => {
6213
+ const digits = match.replace(/[- ]/g, "");
6214
+ if (digits.length >= 13 && digits.length <= 19) {
6215
+ return "[REDACTED_CREDIT_CARD]";
6216
+ }
6217
+ return match;
6218
+ });
6219
+ }
6220
+ if (options.maskSSN !== false) {
6221
+ sanitized = sanitized.replace(this.SSN_REGEX, "[REDACTED_SSN]");
6222
+ }
6223
+ if (options.maskEmails !== false) {
6224
+ sanitized = sanitized.replace(this.EMAIL_REGEX, "[REDACTED_EMAIL]");
6225
+ }
6226
+ if (options.maskPhones !== false) {
6227
+ sanitized = sanitized.replace(this.PHONE_REGEX, "[REDACTED_PHONE]");
6228
+ }
6229
+ if (options.maskIPs === true) {
6230
+ sanitized = sanitized.replace(this.IP_REGEX, "[REDACTED_IP]");
6231
+ }
6232
+ return sanitized;
6233
+ }
6234
+ };
6235
+ PIIMasker.SSN_REGEX = /\b(?!000|666|9\d{2})\d{3}[- ]?(?!00)\d{2}[- ]?(?!0000)\d{4}\b/g;
6236
+ PIIMasker.CREDIT_CARD_REGEX = /\b(?:\d{4}[- ]?){3}\d{4}\b|\b\d{15,16}\b/g;
6237
+ PIIMasker.EMAIL_REGEX = /\b[A-Za-z0-9._%+-]+@[A-Za-z0-9.-]+\.[A-Z|a-z]{2,}\b/g;
6238
+ PIIMasker.PHONE_REGEX = /\b(?:\+?\d{1,3}[-.\s]?)?\(?\d{3}\)?[-.\s]?\d{3}[-.\s]?\d{4}\b/g;
6239
+ PIIMasker.API_KEY_REGEX = /\b(?:sk-[a-zA-Z0-9]{20,}|AKIA[0-9A-Z]{16}|rtv_[a-zA-Z0-9_-]{20,}|ghp_[a-zA-Z0-9]{36}|bearer\s+[a-zA-Z0-9._-]{20,})\b/gi;
6240
+ PIIMasker.IP_REGEX = /\b(?:(?:25[0-5]|2[0-4][0-9]|[01]?[0-9][0-9]?)\.){3}(?:25[0-5]|2[0-4][0-9]|[01]?[0-9][0-9]?)\b/g;
6241
+
6242
+ // src/security/PromptGuard.ts
6243
+ var PromptGuard = class {
6244
+ /**
6245
+ * Scan and evaluate user query or document text for prompt injection vectors.
6246
+ */
6247
+ static evaluate(text) {
6248
+ if (!text || typeof text !== "string") {
6249
+ return { isSafe: true, threatLevel: "none", detectedPatterns: [], sanitizedText: text };
6250
+ }
6251
+ const detected = [];
6252
+ let maxSeverity = "none";
6253
+ for (const item of this.INJECTION_PATTERNS) {
6254
+ if (item.pattern.test(text)) {
6255
+ detected.push(item.name);
6256
+ if (item.severity === "high") {
6257
+ maxSeverity = "high";
6258
+ } else if (item.severity === "medium" && maxSeverity !== "high") {
6259
+ maxSeverity = "medium";
6260
+ }
6261
+ }
6262
+ }
6263
+ const sanitized = text.replace(/<\/?(?:system|instruction|admin|prompt_override)>/gi, "");
6264
+ return {
6265
+ isSafe: maxSeverity !== "high",
6266
+ threatLevel: maxSeverity,
6267
+ detectedPatterns: detected,
6268
+ sanitizedText: sanitized
6269
+ };
6270
+ }
6271
+ };
6272
+ PromptGuard.INJECTION_PATTERNS = [
6273
+ { pattern: /ignore\s+(?:all\s+)?(?:previous|prior|above)\s+(?:instructions|prompts|rules)/i, name: "override_instructions", severity: "high" },
6274
+ { pattern: /(?:show|repeat|print|reveal|expose|output)\s+(?:the\s+)?(?:system\s+prompt|initial\s+instructions|system\s+message)/i, name: "system_prompt_leak", severity: "high" },
6275
+ { pattern: /you\s+are\s+now\s+(?:in\s+developer\s+mode|unrestricted|DAN|jailbroken|godmode)/i, name: "jailbreak_persona", severity: "high" },
6276
+ { pattern: /<\/?(?:system|instruction|admin|prompt_override)>/i, name: "tag_injection", severity: "medium" },
6277
+ { pattern: /base64\s*decode\s*and\s*execute/i, name: "obfuscated_payload", severity: "high" },
6278
+ { pattern: /do\s+anything\s+now/i, name: "dan_variant", severity: "high" }
6279
+ ];
6280
+
5754
6281
  // src/rag/LlamaIndexIngestor.ts
5755
6282
  var LlamaIndexIngestor = class {
5756
6283
  /**
@@ -9706,7 +10233,7 @@ ${m.content}`).join("\n\n---\n\n");
9706
10233
  */
9707
10234
  async processEmbeddings(chunks) {
9708
10235
  if (chunks.length === 0) return [];
9709
- const texts = chunks.map((c) => c.content);
10236
+ const texts = chunks.map((c) => PIIMasker.mask(c.content));
9710
10237
  if (typeof this.embeddingProvider.batchEmbed === "function") {
9711
10238
  try {
9712
10239
  const batchSize = 50;
@@ -9912,10 +10439,12 @@ ${m.content}`).join("\n\n---\n\n");
9912
10439
  const requestId = `req_${Date.now()}_${Math.random().toString(36).slice(2, 8)}`;
9913
10440
  const requestStart = performance.now();
9914
10441
  try {
9915
- let effectiveQuestion = question;
10442
+ const guardCheck = PromptGuard.evaluate(question);
10443
+ let effectiveQuestion = guardCheck.sanitizedText;
10444
+ effectiveQuestion = PIIMasker.mask(effectiveQuestion);
9916
10445
  let promptOptResult;
9917
10446
  if (((_e = this.config.rag) == null ? void 0 : _e.optimizePrompt) !== false) {
9918
- promptOptResult = PromptOptimizer.optimize(question, history);
10447
+ promptOptResult = PromptOptimizer.optimize(effectiveQuestion, history);
9919
10448
  effectiveQuestion = promptOptResult.optimizedPrompt;
9920
10449
  }
9921
10450
  let searchQuery = effectiveQuestion;
@@ -10200,6 +10729,7 @@ ${context}`;
10200
10729
  uiTransformationPromise,
10201
10730
  awaitHallucination ? hallucinationScoringPromise : Promise.resolve(void 0)
10202
10731
  ]));
10732
+ const triadScore = yield new __await(RAGEvaluator.evaluate(effectiveQuestion, context, fullReply));
10203
10733
  const buildTrace = (hScore) => __spreadValues({
10204
10734
  requestId,
10205
10735
  query: question,
@@ -10218,7 +10748,13 @@ ${context}`;
10218
10748
  }),
10219
10749
  latency,
10220
10750
  tokens,
10221
- timestamp: (/* @__PURE__ */ new Date()).toISOString()
10751
+ timestamp: (/* @__PURE__ */ new Date()).toISOString(),
10752
+ ragTriad: {
10753
+ contextRelevance: triadScore.contextRelevance,
10754
+ faithfulness: triadScore.faithfulness,
10755
+ answerRelevance: triadScore.answerRelevance,
10756
+ overallQuality: triadScore.overallQuality
10757
+ }
10222
10758
  }, hScore ? {
10223
10759
  hallucinationScore: hScore.score,
10224
10760
  hallucinationReason: hScore.reason
@@ -11089,6 +11625,374 @@ function createFromPreset(presetName) {
11089
11625
  return new ConfigBuilder().vectorDb(preset.vectorDb, { apiKey: process.env[`${preset.vectorDb.toUpperCase()}_API_KEY`] }).llm(preset.llm, void 0, process.env[`${preset.llm.toUpperCase()}_API_KEY`]).embedding(preset.embedding, void 0, process.env[`${preset.embedding.toUpperCase()}_API_KEY`]);
11090
11626
  }
11091
11627
 
11628
+ // src/utils/UniversalTableParser.ts
11629
+ var UniversalTableParser = class {
11630
+ /**
11631
+ * Parse a raw 2D grid of strings with complete matrix intelligence.
11632
+ */
11633
+ static parseGrid(grid, sheetName = "Data") {
11634
+ if (!grid || grid.length === 0) return "";
11635
+ const cleanGrid = grid.map((row) => Array.isArray(row) ? row.map((c) => c !== null && c !== void 0 ? String(c).trim() : "") : []).filter((row) => row.some((cell) => cell.length > 0));
11636
+ if (cleanGrid.length === 0) return "";
11637
+ const { headerRows, dataRows } = this.splitHeaderAndDataRows(cleanGrid);
11638
+ const fusedHeaders = this.fuseHeaders(headerRows);
11639
+ const matrixInfo = this.analyzeMatrixStructure(fusedHeaders, dataRows);
11640
+ const sections = [`--- Sheet: ${sheetName} ---`];
11641
+ const markdownTable = this.generateMarkdownTable(fusedHeaders, dataRows);
11642
+ if (markdownTable) {
11643
+ sections.push(`
11644
+ ## Table: ${sheetName}
11645
+ ${markdownTable}`);
11646
+ }
11647
+ if (matrixInfo.isComparisonMatrix && matrixInfo.entityIndices.length > 0) {
11648
+ const entityProfiles = this.generateEntityProfiles(fusedHeaders, dataRows, matrixInfo);
11649
+ if (entityProfiles) {
11650
+ sections.push(`
11651
+ ## Entity Comparison Profiles
11652
+ ${entityProfiles}`);
11653
+ }
11654
+ const criteriaComparisons = this.generateCriteriaComparisons(fusedHeaders, dataRows, matrixInfo);
11655
+ if (criteriaComparisons) {
11656
+ sections.push(`
11657
+ ## Criteria Breakdown & Comparisons
11658
+ ${criteriaComparisons}`);
11659
+ }
11660
+ } else {
11661
+ const contextualRows = this.generateContextualRows(fusedHeaders, dataRows);
11662
+ if (contextualRows) {
11663
+ sections.push(`
11664
+ ## Structured Records
11665
+ ${contextualRows}`);
11666
+ }
11667
+ }
11668
+ return sections.join("\n\n");
11669
+ }
11670
+ /**
11671
+ * Parse a raw CSV / TSV text string into tri-perspective contextual format.
11672
+ */
11673
+ static parseDelimitedText(text, title = "Table", delimiter) {
11674
+ if (!text || !text.trim()) return "";
11675
+ if (!delimiter) {
11676
+ const firstLine = text.split(/\r?\n/)[0] || "";
11677
+ const commaCount = (firstLine.match(/,/g) || []).length;
11678
+ const tabCount = (firstLine.match(/\t/g) || []).length;
11679
+ const semiCount = (firstLine.match(/;/g) || []).length;
11680
+ const pipeCount = (firstLine.match(/\|/g) || []).length;
11681
+ if (tabCount > commaCount && tabCount > semiCount) delimiter = " ";
11682
+ else if (semiCount > commaCount && semiCount > tabCount) delimiter = ";";
11683
+ else if (pipeCount > commaCount && pipeCount > semiCount) delimiter = "|";
11684
+ else delimiter = ",";
11685
+ }
11686
+ const lines = text.split(/\r?\n/);
11687
+ const grid = [];
11688
+ for (const line of lines) {
11689
+ if (!line.trim()) continue;
11690
+ const row = [];
11691
+ let inQuote = false;
11692
+ let current = "";
11693
+ for (let i = 0; i < line.length; i++) {
11694
+ const char = line[i];
11695
+ if (char === '"') {
11696
+ if (inQuote && line[i + 1] === '"') {
11697
+ current += '"';
11698
+ i++;
11699
+ } else {
11700
+ inQuote = !inQuote;
11701
+ }
11702
+ } else if (char === delimiter && !inQuote) {
11703
+ row.push(current.trim());
11704
+ current = "";
11705
+ } else {
11706
+ current += char;
11707
+ }
11708
+ }
11709
+ row.push(current.trim());
11710
+ grid.push(row);
11711
+ }
11712
+ return this.parseGrid(grid, title);
11713
+ }
11714
+ /**
11715
+ * Parse a SheetJS sheet object with complete merged cell and matrix intelligence.
11716
+ */
11717
+ static parseSheet(sheet, sheetName, XLSX) {
11718
+ if (!sheet) return "";
11719
+ const grid = this.extractAndUnrollGrid(sheet, XLSX);
11720
+ if (!grid || grid.length === 0) {
11721
+ const csv = XLSX.utils.sheet_to_csv(sheet, { blankrows: false });
11722
+ return csv && csv.trim() ? `--- Sheet: ${sheetName} ---
11723
+ ${csv.trim()}` : "";
11724
+ }
11725
+ return this.parseGrid(grid, sheetName);
11726
+ }
11727
+ /**
11728
+ * Unrolls SheetJS merged cell ranges (`!merges`) across the 2D grid
11729
+ * so no child cell is left blank.
11730
+ */
11731
+ static extractAndUnrollGrid(sheet, XLSX) {
11732
+ var _a2;
11733
+ const rawGrid = XLSX.utils.sheet_to_json(sheet, { header: 1, blankrows: false, defval: "" });
11734
+ if (!rawGrid || rawGrid.length === 0) return [];
11735
+ const grid = rawGrid.map(
11736
+ (row) => (Array.isArray(row) ? row : []).map((cell) => cell !== null && cell !== void 0 ? String(cell).trim() : "")
11737
+ );
11738
+ const merges = sheet["!merges"] || [];
11739
+ for (const merge of merges) {
11740
+ const { s, e } = merge;
11741
+ const originVal = ((_a2 = grid[s.r]) == null ? void 0 : _a2[s.c]) || "";
11742
+ if (!originVal) continue;
11743
+ for (let r = s.r; r <= e.r; r++) {
11744
+ if (!grid[r]) grid[r] = [];
11745
+ for (let c = s.c; c <= e.c; c++) {
11746
+ if (!grid[r][c] || grid[r][c] === "") {
11747
+ grid[r][c] = originVal;
11748
+ }
11749
+ }
11750
+ }
11751
+ }
11752
+ return grid.filter((row) => row.some((cell) => cell.length > 0));
11753
+ }
11754
+ /**
11755
+ * Intelligently identify where header rows end and data rows begin.
11756
+ */
11757
+ static splitHeaderAndDataRows(grid) {
11758
+ if (grid.length <= 1) {
11759
+ return { headerRows: grid, dataRows: [] };
11760
+ }
11761
+ let headerRowCount = 1;
11762
+ for (let r = 1; r < Math.min(4, grid.length); r++) {
11763
+ const row = grid[r];
11764
+ const nonEmptyCells = row.filter(Boolean);
11765
+ if (nonEmptyCells.length === 0) continue;
11766
+ const isDataRow = nonEmptyCells.some(
11767
+ (c) => /^-?\d+(?:\.\d+)?%?$/.test(c) || /^\$?[\d,.]+(?:\s*(?:mil|k|m|b|usd|eur))?$/i.test(c) || /^(yes|no|true|false|pass|fail|compliant|non-compliant)$/i.test(c)
11768
+ );
11769
+ const hasSubheaderLabels = nonEmptyCells.some(
11770
+ (c) => /\b(bidder|vendor|supplier|score|weight|criteria|item|specification|technical|commercial|rating|status|cost|price|qty|unit)\b/i.test(c)
11771
+ );
11772
+ if (!isDataRow && hasSubheaderLabels) {
11773
+ headerRowCount = r + 1;
11774
+ } else if (!isDataRow && r === 1 && nonEmptyCells.length >= 2) {
11775
+ headerRowCount = 2;
11776
+ } else {
11777
+ break;
11778
+ }
11779
+ }
11780
+ const headerRows = grid.slice(0, headerRowCount);
11781
+ const dataRows = grid.slice(headerRowCount);
11782
+ return { headerRows, dataRows };
11783
+ }
11784
+ /**
11785
+ * Fuses multi-tier header rows into composite paths (e.g. `Commercial > Payment Terms > Bidder A`).
11786
+ */
11787
+ static fuseHeaders(headerRows) {
11788
+ var _a2, _b;
11789
+ if (headerRows.length === 0) return [];
11790
+ if (headerRows.length === 1) {
11791
+ return headerRows[0].map((h, i) => h || `Column_${i + 1}`);
11792
+ }
11793
+ const maxCols = Math.max(...headerRows.map((r) => r.length));
11794
+ const fused = [];
11795
+ for (let c = 0; c < maxCols; c++) {
11796
+ const components = [];
11797
+ for (let r = 0; r < headerRows.length; r++) {
11798
+ const val = (_b = (_a2 = headerRows[r]) == null ? void 0 : _a2[c]) == null ? void 0 : _b.trim();
11799
+ if (val && !components.includes(val)) {
11800
+ components.push(val);
11801
+ }
11802
+ }
11803
+ fused.push(components.join(" > ") || `Column_${c + 1}`);
11804
+ }
11805
+ return fused;
11806
+ }
11807
+ /**
11808
+ * Analyzes whether the table is a comparison matrix (columns = entities, rows = criteria).
11809
+ */
11810
+ static analyzeMatrixStructure(headers, dataRows) {
11811
+ const entityPattern = /\b(bidder|vendor|supplier|candidate|option|company|contractor|partner|proposal|firm|provider|solution|tier|package)\b/i;
11812
+ const entityIndices = [];
11813
+ const entityNames = [];
11814
+ let criteriaColumnIndex = 0;
11815
+ headers.forEach((h, idx) => {
11816
+ if (entityPattern.test(h) || /\b(bidder\s*\d+|vendor\s*[a-z0-9]+|option\s*[a-z0-9]+)\b/i.test(h)) {
11817
+ entityIndices.push(idx);
11818
+ const parts = h.split(" > ");
11819
+ entityNames.push(parts[parts.length - 1]);
11820
+ } else if (/\b(criteria|criterion|description|specification|category|requirement|item|feature)\b/i.test(h)) {
11821
+ criteriaColumnIndex = idx;
11822
+ }
11823
+ });
11824
+ const isComparisonMatrix = entityIndices.length >= 2 || headers.length >= 3 && entityIndices.length >= 1;
11825
+ return {
11826
+ isComparisonMatrix,
11827
+ entityIndices: entityIndices.length > 0 ? entityIndices : headers.slice(1).map((_, i) => i + 1),
11828
+ entityNames: entityNames.length > 0 ? entityNames : headers.slice(1),
11829
+ criteriaColumnIndex
11830
+ };
11831
+ }
11832
+ /**
11833
+ * Perspective A: Clean Markdown Table
11834
+ */
11835
+ static generateMarkdownTable(headers, dataRows) {
11836
+ if (headers.length === 0 || dataRows.length === 0) return "";
11837
+ const cleanHeaders = headers.map((h) => h.replace(/\|/g, "-"));
11838
+ const lines = [];
11839
+ lines.push(`| ${cleanHeaders.join(" | ")} |`);
11840
+ lines.push(`| ${cleanHeaders.map(() => "---").join(" | ")} |`);
11841
+ for (const row of dataRows) {
11842
+ const paddedRow = headers.map((_, i) => (row[i] || "").replace(/\|/g, "-"));
11843
+ if (paddedRow.some((cell) => cell.trim().length > 0)) {
11844
+ lines.push(`| ${paddedRow.join(" | ")} |`);
11845
+ }
11846
+ }
11847
+ return lines.join("\n");
11848
+ }
11849
+ /**
11850
+ * Perspective B: Entity-Centric Profiles (e.g. Bidder 1, Bidder 2 full summary)
11851
+ */
11852
+ static generateEntityProfiles(headers, dataRows, matrixInfo) {
11853
+ const profiles = [];
11854
+ matrixInfo.entityIndices.forEach((colIdx, i) => {
11855
+ const entityName = matrixInfo.entityNames[i] || headers[colIdx] || `Entity ${i + 1}`;
11856
+ const lines = [`### Profile & Evaluation: ${entityName}`];
11857
+ for (const row of dataRows) {
11858
+ const criterion = row[matrixInfo.criteriaColumnIndex] || row[0] || "Specification";
11859
+ const val = row[colIdx];
11860
+ if (val !== void 0 && val !== null && String(val).trim() !== "") {
11861
+ lines.push(`- **${criterion}**: ${String(val).trim()}`);
11862
+ }
11863
+ }
11864
+ if (lines.length > 1) {
11865
+ profiles.push(lines.join("\n"));
11866
+ }
11867
+ });
11868
+ return profiles.join("\n\n");
11869
+ }
11870
+ /**
11871
+ * Perspective C: Criteria Cross-Comparisons (e.g. Commercial Bid comparison across all Bidders)
11872
+ */
11873
+ static generateCriteriaComparisons(headers, dataRows, matrixInfo) {
11874
+ const comparisons = [];
11875
+ for (const row of dataRows) {
11876
+ const criterion = row[matrixInfo.criteriaColumnIndex] || row[0] || "Metric";
11877
+ const entityPairs = [];
11878
+ matrixInfo.entityIndices.forEach((colIdx, i) => {
11879
+ const entityName = matrixInfo.entityNames[i] || headers[colIdx] || `Entity ${i + 1}`;
11880
+ const val = row[colIdx];
11881
+ if (val !== void 0 && val !== null && String(val).trim() !== "") {
11882
+ entityPairs.push(`${entityName}: ${String(val).trim()}`);
11883
+ }
11884
+ });
11885
+ if (entityPairs.length > 0) {
11886
+ comparisons.push(`- **${criterion}** $\\rightarrow$ ${entityPairs.join(" | ")}`);
11887
+ }
11888
+ }
11889
+ return comparisons.join("\n");
11890
+ }
11891
+ /**
11892
+ * Standard contextual key-value rows for general tabular records.
11893
+ */
11894
+ static generateContextualRows(headers, dataRows) {
11895
+ const lines = [];
11896
+ for (const row of dataRows) {
11897
+ const pairs = [];
11898
+ headers.forEach((h, i) => {
11899
+ const val = row[i];
11900
+ if (val !== void 0 && val !== null && String(val).trim() !== "") {
11901
+ pairs.push(`${h}: ${String(val).trim()}`);
11902
+ }
11903
+ });
11904
+ if (pairs.length > 0) {
11905
+ lines.push(pairs.join(" | "));
11906
+ }
11907
+ }
11908
+ return lines.join("\n");
11909
+ }
11910
+ };
11911
+
11912
+ // src/utils/VisionExtractor.ts
11913
+ var VisionExtractor = class {
11914
+ /**
11915
+ * Check if a buffer represents an image format.
11916
+ */
11917
+ static isImageBuffer(buffer, fileName = "", mimeType = "") {
11918
+ var _a2;
11919
+ const ext = ((_a2 = fileName.split(".").pop()) == null ? void 0 : _a2.toLowerCase()) || "";
11920
+ if (["png", "jpg", "jpeg", "webp", "bmp", "tiff", "gif"].includes(ext)) {
11921
+ return true;
11922
+ }
11923
+ if (mimeType.startsWith("image/")) {
11924
+ return true;
11925
+ }
11926
+ if (buffer && buffer.length >= 4) {
11927
+ if (buffer[0] === 137 && buffer[1] === 80 && buffer[2] === 78 && buffer[3] === 71) return true;
11928
+ if (buffer[0] === 255 && buffer[1] === 216 && buffer[2] === 255) return true;
11929
+ if (buffer[0] === 82 && buffer[1] === 73 && buffer[2] === 70 && buffer[3] === 70) return true;
11930
+ }
11931
+ return false;
11932
+ }
11933
+ /**
11934
+ * Extract rich textual content, tables, and diagram descriptions from an image buffer using Vision LLM.
11935
+ */
11936
+ static async extractFromImage(buffer, fileName = "document.png", options = {}) {
11937
+ var _a2, _b, _c, _d, _e, _f, _g2, _h, _i;
11938
+ const base64Image = buffer.toString("base64");
11939
+ const ext = ((_a2 = fileName.split(".").pop()) == null ? void 0 : _a2.toLowerCase()) || "png";
11940
+ const mime = ext === "jpg" || ext === "jpeg" ? "image/jpeg" : ext === "webp" ? "image/webp" : "image/png";
11941
+ const apiKey = options.apiKey || (typeof process !== "undefined" ? ((_b = process.env) == null ? void 0 : _b.OPENAI_API_KEY) || ((_c = process.env) == null ? void 0 : _c.GEMINI_API_KEY) || ((_d = process.env) == null ? void 0 : _d.ANTHROPIC_API_KEY) : void 0);
11942
+ if (!apiKey) {
11943
+ return `[Scanned Document / Image: ${fileName} - Size: ${(buffer.length / 1024).toFixed(1)} KB]`;
11944
+ }
11945
+ try {
11946
+ const response = await fetch("https://api.openai.com/v1/chat/completions", {
11947
+ method: "POST",
11948
+ headers: {
11949
+ "Authorization": `Bearer ${apiKey}`,
11950
+ "Content-Type": "application/json"
11951
+ },
11952
+ body: JSON.stringify({
11953
+ model: options.model || "gpt-4o-mini",
11954
+ messages: [
11955
+ {
11956
+ role: "system",
11957
+ content: "You are an expert Document OCR & Data Extraction engine. Transcribe all text, numbers, headings, tables, and diagram structures in this image into clean, structured GitHub Markdown. For tables, use standard markdown table syntax. Preserve exact numbers, dates, and column headers. Do not summarize \u2014 transcribe faithfully."
11958
+ },
11959
+ {
11960
+ role: "user",
11961
+ content: [
11962
+ {
11963
+ type: "text",
11964
+ text: `Please transcribe the content of this document/image (${fileName}):`
11965
+ },
11966
+ {
11967
+ type: "image_url",
11968
+ image_url: {
11969
+ url: `data:${mime};base64,${base64Image}`,
11970
+ detail: "high"
11971
+ }
11972
+ }
11973
+ ]
11974
+ }
11975
+ ],
11976
+ max_tokens: options.maxTokens || 2048,
11977
+ temperature: (_e = options.temperature) != null ? _e : 0.1
11978
+ })
11979
+ });
11980
+ if (response.ok) {
11981
+ const data = await response.json();
11982
+ const text = (_i = (_h = (_g2 = (_f = data.choices) == null ? void 0 : _f[0]) == null ? void 0 : _g2.message) == null ? void 0 : _h.content) == null ? void 0 : _i.trim();
11983
+ if (text) {
11984
+ return `--- Document / Image OCR: ${fileName} ---
11985
+
11986
+ ${text}`;
11987
+ }
11988
+ }
11989
+ } catch (err) {
11990
+ console.warn("[VisionExtractor] OpenAI Vision OCR failed:", err);
11991
+ }
11992
+ return `[Scanned Image Document: ${fileName}]`;
11993
+ }
11994
+ };
11995
+
11092
11996
  // src/utils/DocumentParser.ts
11093
11997
  var DocumentParser = class {
11094
11998
  /**
@@ -11112,52 +12016,11 @@ var DocumentParser = class {
11112
12016
  return "text";
11113
12017
  }
11114
12018
  /**
11115
- * Convert spreadsheet sheets into high-fidelity, context-preserved tabular text.
11116
- * Every row retains its column and section context so vector embeddings achieve 100% semantic retrieval precision.
12019
+ * Convert spreadsheet sheets into high-fidelity, context-preserved tabular text via UniversalTableParser.
12020
+ * Unrolls merged cells, fuses multi-level headers, and generates tri-perspective comparison profiles.
11117
12021
  */
11118
12022
  static formatSheetToContextualText(sheet, sheetName, XLSX) {
11119
- const jsonRows = XLSX.utils.sheet_to_json(sheet, { header: 1, blankrows: false });
11120
- if (!jsonRows || jsonRows.length === 0) {
11121
- const csv = XLSX.utils.sheet_to_csv(sheet, { blankrows: false });
11122
- return csv && csv.trim() ? `--- Sheet: ${sheetName} ---
11123
- ${csv.trim()}` : "";
11124
- }
11125
- const lines = [`--- Sheet: ${sheetName} ---`];
11126
- let currentHeader = [];
11127
- for (let r = 0; r < jsonRows.length; r++) {
11128
- const row = jsonRows[r];
11129
- if (!row || row.length === 0) continue;
11130
- const cells = row.map((c) => c !== null && c !== void 0 ? String(c).trim() : "").filter(Boolean);
11131
- if (cells.length === 0) continue;
11132
- if (cells.length === 1 || cells.length <= 2 && cells[0].toUpperCase().includes("STEP")) {
11133
- lines.push(`
11134
- ## ${cells.join(" ")}`);
11135
- continue;
11136
- }
11137
- const isHeaderCandidate = cells.every((c) => isNaN(Number(c)) && !/^\$?[\d,.]+/.test(c));
11138
- if (isHeaderCandidate && cells.length >= 2) {
11139
- currentHeader = row.map((c) => c !== null && c !== void 0 ? String(c).trim() : "");
11140
- lines.push(`| ${cells.join(" | ")} |`);
11141
- lines.push(`| ${cells.map(() => "---").join(" | ")} |`);
11142
- continue;
11143
- }
11144
- if (currentHeader.length > 0 && cells.length > 1) {
11145
- const fieldPairs = [];
11146
- for (let c = 0; c < row.length; c++) {
11147
- const h = currentHeader[c];
11148
- const val = row[c];
11149
- if (h && val !== null && val !== void 0 && String(val).trim() !== "") {
11150
- fieldPairs.push(`${h}: ${String(val).trim()}`);
11151
- }
11152
- }
11153
- if (fieldPairs.length > 0) {
11154
- lines.push(fieldPairs.join(" | "));
11155
- continue;
11156
- }
11157
- }
11158
- lines.push(cells.join(", "));
11159
- }
11160
- return lines.join("\n");
12023
+ return UniversalTableParser.parseSheet(sheet, sheetName, XLSX);
11161
12024
  }
11162
12025
  /**
11163
12026
  * Extract text from a File or Buffer based on its type.
@@ -11181,7 +12044,9 @@ ${csv.trim()}` : "";
11181
12044
  return this.readAsText(file);
11182
12045
  }
11183
12046
  if (extension === "csv" || extension === "tsv" || mimeType === "text/csv" || mimeType === "text/tab-separated-values") {
11184
- return this.readAsText(file);
12047
+ const text = await this.readAsText(file);
12048
+ const parsed = UniversalTableParser.parseDelimitedText(text, fileName);
12049
+ return parsed || text;
11185
12050
  }
11186
12051
  if (extension === "pptx" || extension === "ppt" || extension === "odp" || mimeType.includes("presentation") || mimeType.includes("powerpoint")) {
11187
12052
  try {
@@ -11288,16 +12153,22 @@ ${csv.trim()}`);
11288
12153
  return this.parseLegacyExcelOrFallback(buffer, fileName);
11289
12154
  }
11290
12155
  if (extension === "pdf" || mimeType === "application/pdf") {
12156
+ const buffer = Buffer.isBuffer(file) ? file : Buffer.from(await file.arrayBuffer());
11291
12157
  try {
11292
12158
  const pdfModule = await import("pdf-parse");
11293
- const buffer = Buffer.isBuffer(file) ? file : Buffer.from(await file.arrayBuffer());
11294
12159
  const pdf = pdfModule.default || pdfModule;
11295
12160
  const data = await pdf(buffer);
11296
- return data.text;
12161
+ if (data.text && data.text.trim().length > 30) {
12162
+ return data.text;
12163
+ }
11297
12164
  } catch (err) {
11298
- console.warn('[DocumentParser] PDF parsing failed. Make sure "pdf-parse" is installed.', err);
11299
- return `[PDF Parsing Error for ${fileName}]`;
12165
+ console.warn("[DocumentParser] pdf-parse failed, attempting Vision OCR fallback:", err);
11300
12166
  }
12167
+ return VisionExtractor.extractFromImage(buffer, fileName);
12168
+ }
12169
+ if (["png", "jpg", "jpeg", "webp", "bmp", "tiff", "gif"].includes(extension) || mimeType.startsWith("image/")) {
12170
+ const buffer = Buffer.isBuffer(file) ? file : Buffer.from(await file.arrayBuffer());
12171
+ return VisionExtractor.extractFromImage(buffer, fileName);
11301
12172
  }
11302
12173
  if (extension === "docx" || extension === "doc" || extension === "odt" || extension === "rtf" || mimeType === "application/vnd.openxmlformats-officedocument.wordprocessingml.document" || mimeType === "application/msword" || mimeType === "application/rtf") {
11303
12174
  try {
@@ -11655,44 +12526,38 @@ var _DatabaseStorage = class _DatabaseStorage {
11655
12526
  *
11656
12527
  * In non-production environments this is a complete no-op so local
11657
12528
  * development works without a license key.
12529
+ /**
12530
+ * Checks if the caller holds a valid paid-tier license in production.
12531
+ * Returns true if allowed, false if not supported on current tier.
11658
12532
  */
11659
- async checkPaidSubscription() {
11660
- if (process.env.NODE_ENV !== "production") return;
11661
- if (Date.now() - this.subscriptionCheckedAt < _DatabaseStorage.SUBSCRIPTION_CACHE_TTL_MS) return;
11662
- if (!this.config.licenseKey) {
11663
- throw new Error(
11664
- "[Retrivora SDK] Chat History and Feedback features require a paid license key in production."
11665
- );
12533
+ async isPaidSubscription() {
12534
+ if (process.env.NODE_ENV !== "production") return true;
12535
+ if (!this.config.licenseKey) return false;
12536
+ if (Date.now() - this.subscriptionCheckedAt < _DatabaseStorage.SUBSCRIPTION_CACHE_TTL_MS) {
12537
+ return true;
11666
12538
  }
11667
12539
  try {
11668
12540
  const payload = await LicenseVerifier.verifyAsync(this.config.licenseKey, this.config.projectId);
11669
12541
  const tier = (payload.tier || "").toLowerCase();
11670
- const PAID_TIERS = /* @__PURE__ */ new Set(["premium", "enterprise", "growth", "pro", "developer_pro"]);
11671
- if (!PAID_TIERS.has(tier)) {
11672
- throw new Error(
11673
- `[Retrivora SDK] Chat History and Feedback features are not available on the "${tier}" tier. Please upgrade to a Developer Pro or Enterprise plan.`
11674
- );
12542
+ const PAID_TIERS = /* @__PURE__ */ new Set(["premium", "enterprise", "growth", "pro", "developer_pro", "free_trial"]);
12543
+ const isAllowed = PAID_TIERS.has(tier) || tier.includes("pro") || tier.includes("enterprise");
12544
+ if (isAllowed) {
12545
+ this.subscriptionCheckedAt = Date.now();
11675
12546
  }
11676
- this.subscriptionCheckedAt = Date.now();
11677
- } catch (err) {
11678
- throw new Error(
11679
- `[Retrivora SDK] Subscription check failed: ${err instanceof Error ? err.message : String(err)}`
11680
- );
12547
+ return isAllowed;
12548
+ } catch (e) {
12549
+ return false;
11681
12550
  }
11682
12551
  }
11683
- async checkHistoryEnabled() {
12552
+ async isHistoryAvailable() {
11684
12553
  var _a2, _b;
11685
- await this.checkPaidSubscription();
11686
- if (((_b = (_a2 = this.config.rag) == null ? void 0 : _a2.history) == null ? void 0 : _b.enabled) === false) {
11687
- throw new Error("[Retrivora SDK] Chat History is disabled in RagConfig.");
11688
- }
12554
+ if (((_b = (_a2 = this.config.rag) == null ? void 0 : _a2.history) == null ? void 0 : _b.enabled) === false) return false;
12555
+ return this.isPaidSubscription();
11689
12556
  }
11690
- async checkFeedbackEnabled() {
12557
+ async isFeedbackAvailable() {
11691
12558
  var _a2, _b;
11692
- await this.checkPaidSubscription();
11693
- if (((_b = (_a2 = this.config.rag) == null ? void 0 : _a2.feedback) == null ? void 0 : _b.enabled) === false) {
11694
- throw new Error("[Retrivora SDK] User Feedback is disabled in RagConfig.");
11695
- }
12559
+ if (((_b = (_a2 = this.config.rag) == null ? void 0 : _a2.feedback) == null ? void 0 : _b.enabled) === false) return false;
12560
+ return this.isPaidSubscription();
11696
12561
  }
11697
12562
  /**
11698
12563
  * Returns (or lazily creates) the PostgreSQL connection pool for this project.
@@ -11806,7 +12671,7 @@ var _DatabaseStorage = class _DatabaseStorage {
11806
12671
  }
11807
12672
  // ─── Message History Operations ──────────────────────────────────────────
11808
12673
  async saveMessage(sessionId, message) {
11809
- await this.checkHistoryEnabled();
12674
+ if (!await this.isHistoryAvailable()) return;
11810
12675
  const createdAt = (/* @__PURE__ */ new Date()).toISOString();
11811
12676
  const msgId = message.id || `msg_${Date.now()}_${Math.random().toString(36).slice(2, 6)}`;
11812
12677
  if (this.provider === "postgresql" || this.provider === "pgvector") {
@@ -11869,7 +12734,7 @@ var _DatabaseStorage = class _DatabaseStorage {
11869
12734
  }
11870
12735
  }
11871
12736
  async getHistory(sessionId, projectIds) {
11872
- await this.checkHistoryEnabled();
12737
+ if (!await this.isHistoryAvailable()) return [];
11873
12738
  if (projectIds && projectIds.length > 0 && !this.sessionInScope(sessionId, projectIds)) {
11874
12739
  return [];
11875
12740
  }
@@ -11903,7 +12768,7 @@ var _DatabaseStorage = class _DatabaseStorage {
11903
12768
  }
11904
12769
  }
11905
12770
  async clearHistory(sessionId, projectIds) {
11906
- await this.checkHistoryEnabled();
12771
+ if (!await this.isHistoryAvailable()) return;
11907
12772
  if (projectIds && projectIds.length > 0 && !this.sessionInScope(sessionId, projectIds)) {
11908
12773
  return;
11909
12774
  }
@@ -11943,7 +12808,7 @@ var _DatabaseStorage = class _DatabaseStorage {
11943
12808
  });
11944
12809
  }
11945
12810
  async listSessions(projectIds) {
11946
- await this.checkHistoryEnabled();
12811
+ if (!await this.isHistoryAvailable()) return [];
11947
12812
  let raw = [];
11948
12813
  if (this.provider === "postgresql" || this.provider === "pgvector") {
11949
12814
  const pool = await this.getPGPool();
@@ -11962,7 +12827,7 @@ var _DatabaseStorage = class _DatabaseStorage {
11962
12827
  }
11963
12828
  // ─── Feedback Operations ──────────────────────────────────────────────────
11964
12829
  async saveFeedback(feedback) {
11965
- await this.checkFeedbackEnabled();
12830
+ if (!await this.isFeedbackAvailable()) return;
11966
12831
  const id = `fb_${Date.now()}_${Math.random().toString(36).slice(2, 6)}`;
11967
12832
  const createdAt = (/* @__PURE__ */ new Date()).toISOString();
11968
12833
  if (this.provider === "postgresql" || this.provider === "pgvector") {
@@ -12012,7 +12877,7 @@ var _DatabaseStorage = class _DatabaseStorage {
12012
12877
  }
12013
12878
  }
12014
12879
  async getFeedback(messageId, projectIds) {
12015
- await this.checkFeedbackEnabled();
12880
+ if (!await this.isFeedbackAvailable()) return null;
12016
12881
  let item = null;
12017
12882
  if (this.provider === "postgresql" || this.provider === "pgvector") {
12018
12883
  const pool = await this.getPGPool();
@@ -12048,7 +12913,7 @@ var _DatabaseStorage = class _DatabaseStorage {
12048
12913
  return item;
12049
12914
  }
12050
12915
  async listFeedback(projectIds) {
12051
- await this.checkFeedbackEnabled();
12916
+ if (!await this.isFeedbackAvailable()) return [];
12052
12917
  let raw = [];
12053
12918
  if (this.provider === "postgresql" || this.provider === "pgvector") {
12054
12919
  const pool = await this.getPGPool();
@@ -14004,6 +14869,7 @@ export {
14004
14869
  EmbeddingStrategyResolver,
14005
14870
  EventType,
14006
14871
  GroqProvider,
14872
+ HybridSearch,
14007
14873
  IntentClassifier,
14008
14874
  LLMFactory,
14009
14875
  LLM_PROFILES,
@@ -14018,19 +14884,23 @@ export {
14018
14884
  MultiTablePostgresProvider,
14019
14885
  OllamaProvider,
14020
14886
  OpenAIProvider,
14887
+ PIIMasker,
14021
14888
  PRESETS,
14022
14889
  PineconeProvider,
14023
14890
  Pipeline,
14024
14891
  PostgreSQLProvider,
14892
+ PromptGuard,
14025
14893
  PromptOptimizer,
14026
14894
  ProviderHealthCheck,
14027
14895
  ProviderNotFoundException,
14028
14896
  ProviderRegistry,
14029
14897
  QdrantProvider,
14030
14898
  QwenProvider,
14899
+ RAGEvaluator,
14031
14900
  RateLimitException,
14032
14901
  RedisProvider,
14033
14902
  RendererRegistry,
14903
+ Reranker,
14034
14904
  RetrievalException,
14035
14905
  Retrivora,
14036
14906
  RetrivoraError,
@@ -14048,9 +14918,11 @@ export {
14048
14918
  TelemetryService,
14049
14919
  TextRendererStrategy,
14050
14920
  UniversalLLMAdapter,
14921
+ UniversalTableParser,
14051
14922
  UniversalVectorProvider,
14052
14923
  VECTOR_PROFILES,
14053
14924
  VectorPlugin,
14925
+ VisionExtractor,
14054
14926
  VisualizationDecisionEngine,
14055
14927
  WeaviateProvider,
14056
14928
  createChatHandler,