Add shared library for APIs/chunking

lib/anthropic.js

@@ -0,0 +1,154 @@
+import Anthropic from "@anthropic-ai/sdk";
+
+const GENERATION_MODEL = "claude-opus-4-6";
+const NLI_MODEL = "claude-haiku-4-5";
+
+
+const RESPONSE_SCHEMA = {
+    type: "object",
+    properties: {
+        answer: {
+            type: "string",
+            description:
+                "The answer with inline numeric citations like [1], [2] matching the source document numbers provided",
+        },
+        citedSourceIndices: {
+            type: "array",
+            items: { type: "integer" },
+            description: "The document numbers cited in the answer",
+        },
+        followUpQuestions: {
+            type: "array",
+            items: { type: "string" },
+            description: "Exactly 3 follow-up questions the user might ask",
+        },
+    },
+    required: ["answer", "citedSourceIndices", "followUpQuestions"],
+    additionalProperties: false,
+};
+
+const NLI_SCHEMA = {
+    type: "object",
+    properties: {
+        label: {
+            type: "string",
+            enum: ["supported", "contradicted", "unsupported"],
+            description: "Whether the claim is entailed, contradicted, or neither by the passages",
+        },
+        confidence: {
+            type: "number",
+            description: "Confidence in the label, from 0 to 1",
+        },
+    },
+    required: ["label", "confidence"],
+    additionalProperties: false,
+};
+
+function getClient() {
+    const key = process.env.ANTHROPIC_API_KEY;
+    if (!key) throw new Error("ANTHROPIC_API_KEY is not set");
+    return new Anthropic({ apiKey: key });
+}
+
+function textFromMessage(message) {
+    const block = message.content[0];
+    return block && "text" in block ? block.text : "";
+}
+
+function buildPrompt(query, sources) {
+    const today = new Date().toLocaleDateString("en-US", {
+        weekday: "long",
+        year: "numeric",
+        month: "long",
+        day: "numeric",
+    });
+
+    const sourceBlock = sources
+        .map((s) => `[Source ${s.docIndex}] (${s.name})\n${s.text}`)
+        .join("\n\n---\n\n");
+
+    return `You are a research assistant. Today's date is ${today}. Answer the user's question using ONLY the source documents provided below. Follow these rules strictly:
+
+1. Ground every claim in a specific source document. Cite sources inline using numeric notation like [1], [2], etc., matching the source document numbers below.
+2. If the sources do not contain enough information to answer, say so honestly.
+3. After your answer, suggest exactly 3 follow-up questions the user might ask based on the sources.
+
+--- SOURCE DOCUMENTS ---
+
+${sourceBlock}
+
+--- USER QUESTION ---
+
+${query}`;
+}
+
+function stripInvalidCitations(answer, validIndices) {
+    return answer
+        .replace(/\[(\d+)\]/g, (match, n) => (validIndices.has(Number(n)) ? match : ""))
+        .replace(/[ \t]{2,}/g, " ")
+        .trim();
+}
+
+export async function generate(query, sources) {
+    const client = getClient();
+    const message = await client.messages.create({
+        model: GENERATION_MODEL,
+        max_tokens: 2048,
+        messages: [{ role: "user", content: buildPrompt(query, sources) }],
+        output_config: {
+            format: {
+                type: "json_schema",
+                schema: RESPONSE_SCHEMA,
+            },
+        },
+    });
+
+    const parsed = JSON.parse(textFromMessage(message) || "{}");
+    const validIndices = new Set(sources.map((s) => s.docIndex));
+    const citedSourceIndices = [
+        ...new Set((parsed.citedSourceIndices || []).filter((idx) => validIndices.has(idx))),
+    ];
+
+    return {
+        answer: stripInvalidCitations(parsed.answer || "", validIndices),
+        citedSourceIndices,
+        followUpQuestions: parsed.followUpQuestions || [],
+    };
+}
+
+export async function judgeEntailment(sentence, chunkTexts) {
+    const client = getClient();
+    const passages = chunkTexts.map((t, i) => `[Passage ${i + 1}]\n${t}`).join("\n\n");
+
+    const prompt = `You are an NLI judge. Decide whether the CLAIM is supported, contradicted, or unsupported by the SOURCE PASSAGES.
+
+- supported: the passages entail the claim
+- contradicted: the passages contradict the claim
+- unsupported: the passages neither support nor contradict the claim
+
+--- CLAIM ---
+${sentence}
+
+--- SOURCE PASSAGES ---
+${passages}`;
+
+    const message = await client.messages.create({
+        model: NLI_MODEL,
+        max_tokens: 256,
+        messages: [{ role: "user", content: prompt }],
+        output_config: {
+            format: {
+                type: "json_schema",
+                schema: NLI_SCHEMA,
+            },
+        },
+    });
+
+    const parsed = JSON.parse(textFromMessage(message) || "{}");
+    const label = ["supported", "contradicted", "unsupported"].includes(parsed.label)
+        ? parsed.label
+        : "unsupported";
+    const confidence = typeof parsed.confidence === "number" ? parsed.confidence : 0;
+
+    return { label, confidence };
+}