@absolutejs/rag 0.22.7 → 0.22.9

package/dist/research/index.js

@@ -101400,6 +101400,28 @@ var init_pdf = __esm(() => {
   };
 });
 
+// src/research/sourceExcerpts.ts
+import { createAITextSource } from "@absolutejs/ai";
+var selectResearchExcerpts = (text, budget, query) => {
+  if (budget <= 0)
+    return [];
+  if (text.length <= budget)
+    return [text];
+  const source = createAITextSource([{ role: "user", content: text }]);
+  const hits = source.search(query);
+  const passages = hits.length ? hits : source.search("");
+  let remaining = budget;
+  const selected = [];
+  for (const passage of passages) {
+    if (!remaining)
+      break;
+    const excerpt = passage.text.slice(0, remaining);
+    selected.push({ start: passage.start, text: excerpt });
+    remaining -= excerpt.length;
+  }
+  return selected.sort((a, b) => a.start - b.start).map((passage) => passage.text);
+};
+
 // src/research/budget.ts
 var createResearchBudgetAdmission = (config) => {
   if (!config.scope.trim())
@@ -120842,7 +120864,7 @@ var readRAGWebsite = async (options) => {
     const page = await readRAGWebpage({
       ...options,
       url: next.url,
-      maxChars: Math.min(maxChars, 12000),
+      maxChars,
       signal: AbortSignal.any([signal, AbortSignal.timeout(30000)])
     });
     pages.push(page);
@@ -120898,24 +120920,27 @@ var readRAGWebsite = async (options) => {
   let remainingDocumentChars = maxChars;
   const documents = sources.flatMap((source) => {
     const page = pages.find((item) => item.finalUrl === source.url && item.text);
-    if (!page || remainingDocumentChars <= 0)
+    if (!page)
       return [];
     const imageLabels = page.evidence.images.filter((image) => image.label).map((image) => `${image.context ? `${image.context}: ` : ""}${image.label}`);
     const fullText = [
-      page.text,
+      page.fullText ?? page.text,
       ...imageLabels.length ? [
         "Image labels supplied by the website (not independently verified customer relationships):",
         ...imageLabels
       ] : []
     ].join(`
 `);
-    const text = fullText.slice(0, remainingDocumentChars);
+    const text = fullText.slice(0, Math.max(0, remainingDocumentChars));
     remainingDocumentChars -= text.length;
     return [
       {
         sourceId: source.id,
         text,
-        truncated: page.truncated || text.length < fullText.length
+        fullText,
+        originalChars: fullText.length,
+        sourceTruncated: page.status !== "ok" || page.fullText === undefined && page.truncated,
+        truncated: page.fullText === undefined && page.truncated || text.length < fullText.length
       }
     ];
   });
@@ -120934,10 +120959,11 @@ ${page.text}`).join(`
     sources,
     citationRequirements: "The final answer must contain clickable Markdown source links, not just source names or bare domain mentions. Reuse sources[].citation beside the factual paragraph, list or table row it supports. Cite the actual supporting page, not the homepage for everything. Every factual section needs its supporting links. Do not invent sources or use unread links as evidence. Before responding, check that company facts, service descriptions, customer examples and numerical claims have supporting links; retrieve or omit unsupported claims. A list of large customers does not prove smaller customers are excluded. Preserve the brands actually named in the evidence; do not append parent companies, ownership relationships, current-account status, market rankings or rebrand history from memory. Label reasoned inferences as such. A verified brand change needs at most one short, cited sentence unless the user asks for its history.",
     ...first,
+    fullText: undefined,
     status: readable.length ? incomplete ? "partial" : "ok" : "error",
     text: evidenceText.slice(0, maxChars),
     truncated: evidenceText.length > maxChars || pages.some((page) => page.truncated),
-    pages: pages.map(({ text: _text, ...page }) => page),
+    pages: pages.map(({ text: _text, fullText: _fullText, ...page }) => page),
     captionEvidence,
     coverage: {
       pagesRead: pages.length,
@@ -120948,10 +120974,10 @@ ${page.text}`).join(`
       deadlineReached: signal.aborted,
       exhaustive: false,
       stopReason: signal.aborted ? "deadline" : unvisited.length ? "page_limit" : "links_exhausted",
-      incompleteReads: pages.filter((page) => page.status !== "ok" || page.truncated).map((page) => ({
+      incompleteReads: pages.filter((page) => page.status !== "ok" || page.fullText === undefined && page.truncated).map((page) => ({
         url: page.finalUrl,
         status: page.status,
-        truncated: page.truncated,
+        truncated: page.fullText === undefined && page.truncated,
         error: page.error
       }))
     },
@@ -121066,12 +121092,15 @@ var createWebpageReaderClient = (options) => {
     if (![evidence.links, evidence.images, evidence.media].every(Array.isArray))
       throw new Error("Invalid page evidence");
     const limit = Math.max(1000, Math.min(input.maxChars ?? 24000, 1e5));
+    const fullText = typeof document2.fullText === "string" ? document2.fullText : document2.text;
     return {
       status: result.status,
       url: input.url,
       finalUrl: result.finalUrl,
       title: typeof result.title === "string" ? result.title : null,
-      text: document2.text.slice(0, limit),
+      text: fullText.slice(0, limit),
+      fullText,
+      originalChars: fullText.length,
       method: result.method,
       truncated: result.truncated === true || document2.truncated === true || document2.text.length > limit,
       fetchedAt: result.fetchedAt,
@@ -121114,6 +121143,8 @@ var readRAGWebpage = async (options) => {
     finalUrl,
     title,
     text: text.slice(0, maxChars),
+    fullText: text,
+    originalChars: text.length,
     method,
     truncated: text.length > maxChars,
     attempts,
@@ -121453,6 +121484,7 @@ ${prompt}`,
     const requiredPhrases = input.requiredPhrases ?? (input.queryMode === "task" ? [] : [...input.query.matchAll(/"([^"\n]+)"/gu)].map((match) => match[1]));
     const urls = new Set;
     const queries = new Set;
+    const discoveredLinks = new Map;
     let evidenceChars = 0;
     const sourceText = new Map;
     const rebalanceEvidence = () => {
@@ -121464,8 +121496,8 @@ ${prompt}`,
         const text = sourceText.get(source.url);
         const allowance = Math.floor(remaining * weight(source) / remainingWeight);
         remainingWeight -= weight(source);
-        source.excerpts = [text.slice(0, allowance)];
-        remaining -= source.excerpts[0].length;
+        source.excerpts = selectResearchExcerpts(text, allowance, [input.query, ...input.queries ?? []].join(" "));
+        remaining -= source.excerpts.reduce((sum, excerpt) => sum + excerpt.length, 0);
       }
       evidenceChars = limits.evidenceChars - remaining;
     };
@@ -121524,7 +121556,7 @@ ${prompt}`,
     const doRead = async (url) => {
       if (urls.has(url) || urls.size >= limits.reads)
         return;
-      if (!result.searches.some((search) => search.sources.some((source) => source.url === url)))
+      if (!discoveredLinks.has(url) && !result.searches.some((search) => search.sources.some((source) => source.url === url)))
         return;
       urls.add(url);
       const page = await operate("read", () => (config.reader ?? readRAGWebpage)({
@@ -121537,9 +121569,19 @@ ${prompt}`,
       if (page.status !== "ok")
         result.limitations.push(`Page read ${page.status}: ${url}`);
       if (page.text && page.status !== "error") {
+        for (const link of page.evidence.links) {
+          if (discoveredLinks.size >= 64)
+            break;
+          try {
+            const target = validatePublicWebUrl(new URL(link.url, page.finalUrl).href);
+            target.hash = "";
+            if (!discoveredLinks.has(target.href))
+              discoveredLinks.set(target.href, { url: target.href, label: link.label.slice(0, 200), discoveredOn: page.finalUrl });
+          } catch {}
+        }
         const prior = result.sources.find((source) => source.url === url);
         if (prior) {
-          sourceText.set(prior.url, page.text.slice(0, limits.evidenceChars));
+          sourceText.set(prior.url, page.fullText ?? page.text);
           prior.contentFetchedAt = page.fetchedAt;
           prior.metadata = { ...prior.metadata, readFinalUrl: page.finalUrl };
           if (page.finalUrl !== prior.url)
@@ -121575,11 +121617,12 @@ ${prompt}`,
           break;
         let plan;
         try {
-          plan = await generate("plan", PlanSchema, "Select follow-up searches for missing evidence and primary-source URLs to read. For current people or roles, seek the organization\u2019s own leadership/about/team pages; testimonials, package authors, and job listings are not staff identity evidence. For policies, read the full policy or agreement to preserve section scope and exceptions. URLs must come from the supplied search results. done means available evidence addresses the requested schema, not merely that some results exist. Return no actions if further work is not useful.", {
+          plan = await generate("plan", PlanSchema, "Select follow-up searches for missing evidence and primary-source URLs to read. For current people or roles, seek the organization\u2019s own leadership/about/team pages; testimonials, package authors, and job listings are not staff identity evidence. For policies, read the full policy or agreement to preserve section scope and exceptions. URLs must come from the supplied search results or discoveredLinks. Discovered links are unread navigation hints, not evidence; read relevant product, pricing, checkout or policy details before concluding information is absent. done means available evidence addresses the requested schema, not merely that some results exist. Return no actions if further work is not useful.", {
             query: input.query,
             schema: task.schema,
             instructions: task.instructions,
             sources: result.sources,
+            discoveredLinks: [...discoveredLinks.values()].filter((link) => !urls.has(link.url)),
             searched: [...queries],
             read: [...urls],
             remainingSearches: limits.searches - queries.size,
@@ -121651,6 +121694,7 @@ ${prompt}`,
       result.status = result.sources.length ? "partial" : "unavailable";
       return result;
     } finally {
+      result.documents = result.sources.filter((source) => source.contentFetchedAt).map((source) => ({ sourceId: source.id, url: source.url, text: sourceText.get(source.url) ?? "", originalChars: (sourceText.get(source.url) ?? "").length }));
       result.limitations = [...new Set(result.limitations)];
       input.onProgress?.({
         phase: "complete",
@@ -122499,5 +122543,5 @@ export {
   selectResearchMonitorEvents
 };
 
-//# debugId=F48D3A6490A0A2CE64756E2164756E21
+//# debugId=B2B2CCEDC66F56B564756E2164756E21
 //# sourceMappingURL=index.js.map