onbuzz 4.9.13 → 4.10.0

package/src/tools/foundryWebSearchTool.js

@@ -1,273 +1,273 @@
-/**
- * @file tools/foundryWebSearchTool.js
- *
- * Azure AI Foundry web-search tool — wraps the backend's
- * /llm/web-search endpoint (Foundry Responses API + Grounding with
- * Bing Search).
- *
- * What this tool is FOR:
- *   - "Research X and cite your sources" — agent gets a synthesized
- *     answer + URL citations in one round-trip.
- *   - Time-sensitive lookups ("latest version of …", "what happened
- *     this week with …") where the model would otherwise guess from
- *     training data.
- *   - Anywhere you want CITATIONS in the response, not just raw URLs.
- *
- * What this tool is NOT for:
- *   - "Give me 10 raw search result links" — use the `web` tool
- *     (puppeteer-based) for that. The Foundry path doesn't expose
- *     raw URLs; it only returns what the model cited.
- *   - Sites that need login/interaction — Foundry's tool reads
- *     public web pages only. The `web` tool handles authenticated
- *     browsing.
- *
- * Architecture: model-centric. The agent decides locally to call this
- * tool; the tool fires ONE backend HTTP request; the backend fires
- * ONE Foundry HTTP request; Foundry runs the model+search+model loop
- * server-side and returns a single answer. No streaming, no agent
- * state on Foundry's side.
- */
-
-import { BaseTool } from './baseTool.js';
-import TagParser from '../utilities/tagParser.js';
-
-class FoundryWebSearchTool extends BaseTool {
-  constructor(config = {}, logger = null) {
-    super(config, logger);
-    this.id = 'foundry_web_search';
-    this.name = 'Foundry Web Search (Grounded)';
-    this.version = '1.0.0';
-    this.requiresProject = false;
-    this.isAsync = true;
-    // Foundry round-trip is normally 2-5s but multi-search runs can
-    // hit 10s+. Don't penalize the agent with a long enforced wait —
-    // the aiService timeout is the real ceiling.
-    this.builtinDelay = 500;
-
-    // Reference to the AI service holding baseUrl + apiKeyManager.
-    // Set by ToolsRegistry.registerTool via setAIService().
-    this.aiService = null;
-  }
-
-  /**
-   * Injected by the registry so we can reuse the auth + URL plumbing
-   * the image / video tools already use. Without it, the tool throws
-   * a clear "not initialised" error rather than half-running.
-   */
-  setAIService(aiService) {
-    this.aiService = aiService;
-  }
-
-  getDescription() {
-    return `
-Foundry Web Search Tool: Grounded web search via Azure AI Foundry +
-Bing. Returns a synthesized answer with citation URLs. Use this when
-you need CURRENT information from the web AND want sources cited.
-
-USAGE:
-
-\`\`\`json
-{
-  "toolId": "foundry_web_search",
-  "parameters": {
-    "query": "What is the current Node.js LTS version this week?"
-  }
-}
-\`\`\`
-
-Optional parameters:
-  - "instructions": override the default tool-use nudge. Use only when
-    the model is ignoring the search; the default already says
-    "always call the web_search_preview tool".
-  - "model": chat model to orchestrate the search. Defaults to "gpt-4o"
-    (must have Grounding with Bing Search bound in Foundry).
-
-XML form (when JSON isn't convenient):
-
-\`\`\`
-<foundry_web_search>
-  <query>Latest stable Postgres version</query>
-</foundry_web_search>
-\`\`\`
-
-RESPONSE SHAPE:
-  - answer    : string (model's synthesized answer)
-  - citations : array of { url, title, start, end } — the URLs the
-                model cited, with indices into \`answer\` so you can
-                render inline links.
-  - usage     : { inputTokens, outputTokens, webSearches } — number
-                of Bing searches actually invoked. If \`webSearches:0\`
-                the model didn't use the tool (it thought it already
-                knew the answer); rephrase the query with explicit
-                "search the web for ..." wording.
-
-WHEN TO USE:
-  ✓ "Research X and cite sources"
-  ✓ "What's the latest …" (time-sensitive)
-  ✓ "Compare these companies' recent earnings" (factual + current)
-
-WHEN NOT TO USE:
-  ✗ "Give me 10 raw search links" — use the \`web\` tool
-  ✗ "Log in to … and …" — use the \`web\` tool
-  ✗ Questions answerable from training data (waste of $)
-
-COST:
-  Per call: ~$0.007 (Bing) + chat-completion tokens ($0.01–0.05 typical).
-  The model may invoke multiple searches per call — usage.webSearches
-  shows how many fired. Each one bills.
-
-LIMITATIONS:
-  Only the URLs the model CITED come back. The raw Bing results list
-  is hidden. For "show me all 10 results from a search", use the
-  \`web\` tool.
-    `.trim();
-  }
-
-  /**
-   * Parse parameters from tool command content. Supports both
-   * raw-XML form and JSON-in-content form (the latter is the
-   * standard agentic invocation).
-   */
-  parseParameters(content) {
-    try {
-      // Prefer JSON: most agents emit JSON tool calls.
-      try {
-        const trimmed = (content || '').trim();
-        if (trimmed.startsWith('{')) {
-          const parsed = JSON.parse(trimmed);
-          // Agents may pass either {query: ...} OR {parameters: {query: ...}}
-          const inner = parsed.parameters || parsed;
-          return {
-            query: typeof inner.query === 'string' ? inner.query.trim() : null,
-            instructions: typeof inner.instructions === 'string' ? inner.instructions : undefined,
-            model: typeof inner.model === 'string' ? inner.model : undefined,
-          };
-        }
-      } catch { /* fall through to XML */ }
-
-      const queryMatches = TagParser.extractContent(content, 'query');
-      const instructionsMatches = TagParser.extractContent(content, 'instructions');
-      const modelMatches = TagParser.extractContent(content, 'model');
-      return {
-        query: queryMatches.length > 0 ? queryMatches[0].trim() : null,
-        instructions: instructionsMatches.length > 0 ? instructionsMatches[0].trim() : undefined,
-        model: modelMatches.length > 0 ? modelMatches[0].trim() : undefined,
-      };
-    } catch (error) {
-      throw new Error(`Failed to parse foundry_web_search parameters: ${error.message}`);
-    }
-  }
-
-  getSupportedActions() {
-    return ['search'];
-  }
-
-  /**
-   * Execute the search.
-   */
-  async execute(params, context) {
-    if (!this.aiService) {
-      return {
-        success: false,
-        error: 'foundry_web_search not initialised: aiService reference missing',
-        output: 'Internal configuration issue. This tool needs the AI service injected at startup.',
-      };
-    }
-
-    const { query, instructions, model } = params || {};
-    if (!query) {
-      return {
-        success: false,
-        error: 'query is required',
-        output: 'Provide a `query` parameter with the search request.',
-      };
-    }
-
-    try {
-      // Idempotency key derived from agent + content hash; lets a
-      // repeated identical query inside the same turn reuse the
-      // prior billing row (the backend dedups on the same key).
-      const idempotencyKey = this._buildIdempotencyKey(context, query);
-
-      const result = await this.aiService.foundryWebSearch(query, {
-        instructions,
-        model,
-        idempotencyKey,
-        sessionId: context?.sessionId,
-        platformProvided: context?.platformProvided !== false,
-      });
-
-      // Format for agent consumption. The answer goes in `output` so
-      // the agent reads it as the primary result; structured fields
-      // (citations, usage) live alongside for any downstream tool
-      // that wants to act on URLs.
-      const citationsBlock = (result.citations || []).length > 0
-        ? '\n\nSources:\n' + result.citations.map((c, i) =>
-            `  [${i + 1}] ${c.title || '(untitled)'} — ${c.url}`).join('\n')
-        : '\n\n(No sources cited.)';
-
-      const searchCount = result.usage?.webSearches || 0;
-      const meta = ` [searches=${searchCount}, tokens=${result.usage?.inputTokens || 0}in/${result.usage?.outputTokens || 0}out]`;
-
-      return {
-        success: true,
-        output: result.answer + citationsBlock + meta,
-        // Structured fields downstream tools can inspect:
-        answer: result.answer,
-        citations: result.citations,
-        usage: result.usage,
-        model: result.model,
-        // Hint to the agent: if webSearches=0, the model didn't
-        // actually search. Surface this so the agent can decide to
-        // retry with stronger wording.
-        ...(searchCount === 0 ? { warning: 'Model did not invoke the web search tool. Try phrasing the query as "Search the web for: ..." to force tool use.' } : {}),
-      };
-
-    } catch (error) {
-      // 503 = feature flag off / not provisioned. Distinct error path
-      // so the agent can fall back to the `web` tool instead of
-      // retrying.
-      if (error.status === 503) {
-        return {
-          success: false,
-          error: error.message,
-          output: error.message + '\n\nFallback: use the `web` tool for puppeteer-based search instead.',
-          notProvisioned: true,
-        };
-      }
-      return {
-        success: false,
-        error: error.message,
-        output: `Web search failed: ${error.message}`,
-      };
-    }
-  }
-
-  /**
-   * Build an idempotency key per (agentId, query). Stable for repeated
-   * identical queries inside the same turn — the backend's
-   * Transaction.findOne dedup short-circuits the duplicate billing.
-   *
-   * Different turns or different agents → different keys → fresh
-   * search + bill, which is correct (the world moves on between
-   * turns, the user expects a fresh result).
-   *
-   * @private
-   */
-  _buildIdempotencyKey(context, query) {
-    const agentId = context?.agentId || 'unknown';
-    const turn = context?.turnId || context?.messageId || Date.now();
-    // Cheap, deterministic, no-deps hash. Doesn't need to be
-    // cryptographic — the keyspace is per-agent so collision risk
-    // is minimal.
-    let h = 0;
-    const s = `${agentId}|${turn}|${query}`;
-    for (let i = 0; i < s.length; i++) {
-      h = ((h << 5) - h) + s.charCodeAt(i);
-      h |= 0;
-    }
-    return `fws-${agentId}-${(h >>> 0).toString(36)}`;
-  }
-}
-
-export default FoundryWebSearchTool;
+/**
+ * @file tools/foundryWebSearchTool.js
+ *
+ * Azure AI Foundry web-search tool — wraps the backend's
+ * /llm/web-search endpoint (Foundry Responses API + Grounding with
+ * Bing Search).
+ *
+ * What this tool is FOR:
+ *   - "Research X and cite your sources" — agent gets a synthesized
+ *     answer + URL citations in one round-trip.
+ *   - Time-sensitive lookups ("latest version of …", "what happened
+ *     this week with …") where the model would otherwise guess from
+ *     training data.
+ *   - Anywhere you want CITATIONS in the response, not just raw URLs.
+ *
+ * What this tool is NOT for:
+ *   - "Give me 10 raw search result links" — use the `web` tool
+ *     (puppeteer-based) for that. The Foundry path doesn't expose
+ *     raw URLs; it only returns what the model cited.
+ *   - Sites that need login/interaction — Foundry's tool reads
+ *     public web pages only. The `web` tool handles authenticated
+ *     browsing.
+ *
+ * Architecture: model-centric. The agent decides locally to call this
+ * tool; the tool fires ONE backend HTTP request; the backend fires
+ * ONE Foundry HTTP request; Foundry runs the model+search+model loop
+ * server-side and returns a single answer. No streaming, no agent
+ * state on Foundry's side.
+ */
+
+import { BaseTool } from './baseTool.js';
+import TagParser from '../utilities/tagParser.js';
+
+class FoundryWebSearchTool extends BaseTool {
+  constructor(config = {}, logger = null) {
+    super(config, logger);
+    this.id = 'foundry_web_search';
+    this.name = 'Foundry Web Search (Grounded)';
+    this.version = '1.0.0';
+    this.requiresProject = false;
+    this.isAsync = true;
+    // Foundry round-trip is normally 2-5s but multi-search runs can
+    // hit 10s+. Don't penalize the agent with a long enforced wait —
+    // the aiService timeout is the real ceiling.
+    this.builtinDelay = 500;
+
+    // Reference to the AI service holding baseUrl + apiKeyManager.
+    // Set by ToolsRegistry.registerTool via setAIService().
+    this.aiService = null;
+  }
+
+  /**
+   * Injected by the registry so we can reuse the auth + URL plumbing
+   * the image / video tools already use. Without it, the tool throws
+   * a clear "not initialised" error rather than half-running.
+   */
+  setAIService(aiService) {
+    this.aiService = aiService;
+  }
+
+  getDescription() {
+    return `
+Foundry Web Search Tool: Grounded web search via Azure AI Foundry +
+Bing. Returns a synthesized answer with citation URLs. Use this when
+you need CURRENT information from the web AND want sources cited.
+
+USAGE:
+
+\`\`\`json
+{
+  "toolId": "foundry_web_search",
+  "parameters": {
+    "query": "What is the current Node.js LTS version this week?"
+  }
+}
+\`\`\`
+
+Optional parameters:
+  - "instructions": override the default tool-use nudge. Use only when
+    the model is ignoring the search; the default already says
+    "always call the web_search_preview tool".
+  - "model": chat model to orchestrate the search. Defaults to "gpt-4o"
+    (must have Grounding with Bing Search bound in Foundry).
+
+XML form (when JSON isn't convenient):
+
+\`\`\`
+<foundry_web_search>
+  <query>Latest stable Postgres version</query>
+</foundry_web_search>
+\`\`\`
+
+RESPONSE SHAPE:
+  - answer    : string (model's synthesized answer)
+  - citations : array of { url, title, start, end } — the URLs the
+                model cited, with indices into \`answer\` so you can
+                render inline links.
+  - usage     : { inputTokens, outputTokens, webSearches } — number
+                of Bing searches actually invoked. If \`webSearches:0\`
+                the model didn't use the tool (it thought it already
+                knew the answer); rephrase the query with explicit
+                "search the web for ..." wording.
+
+WHEN TO USE:
+  ✓ "Research X and cite sources"
+  ✓ "What's the latest …" (time-sensitive)
+  ✓ "Compare these companies' recent earnings" (factual + current)
+
+WHEN NOT TO USE:
+  ✗ "Give me 10 raw search links" — use the \`web\` tool
+  ✗ "Log in to … and …" — use the \`web\` tool
+  ✗ Questions answerable from training data (waste of $)
+
+COST:
+  Per call: ~$0.007 (Bing) + chat-completion tokens ($0.01–0.05 typical).
+  The model may invoke multiple searches per call — usage.webSearches
+  shows how many fired. Each one bills.
+
+LIMITATIONS:
+  Only the URLs the model CITED come back. The raw Bing results list
+  is hidden. For "show me all 10 results from a search", use the
+  \`web\` tool.
+    `.trim();
+  }
+
+  /**
+   * Parse parameters from tool command content. Supports both
+   * raw-XML form and JSON-in-content form (the latter is the
+   * standard agentic invocation).
+   */
+  parseParameters(content) {
+    try {
+      // Prefer JSON: most agents emit JSON tool calls.
+      try {
+        const trimmed = (content || '').trim();
+        if (trimmed.startsWith('{')) {
+          const parsed = JSON.parse(trimmed);
+          // Agents may pass either {query: ...} OR {parameters: {query: ...}}
+          const inner = parsed.parameters || parsed;
+          return {
+            query: typeof inner.query === 'string' ? inner.query.trim() : null,
+            instructions: typeof inner.instructions === 'string' ? inner.instructions : undefined,
+            model: typeof inner.model === 'string' ? inner.model : undefined,
+          };
+        }
+      } catch { /* fall through to XML */ }
+
+      const queryMatches = TagParser.extractContent(content, 'query');
+      const instructionsMatches = TagParser.extractContent(content, 'instructions');
+      const modelMatches = TagParser.extractContent(content, 'model');
+      return {
+        query: queryMatches.length > 0 ? queryMatches[0].trim() : null,
+        instructions: instructionsMatches.length > 0 ? instructionsMatches[0].trim() : undefined,
+        model: modelMatches.length > 0 ? modelMatches[0].trim() : undefined,
+      };
+    } catch (error) {
+      throw new Error(`Failed to parse foundry_web_search parameters: ${error.message}`, { cause: error });
+    }
+  }
+
+  getSupportedActions() {
+    return ['search'];
+  }
+
+  /**
+   * Execute the search.
+   */
+  async execute(params, context) {
+    if (!this.aiService) {
+      return {
+        success: false,
+        error: 'foundry_web_search not initialised: aiService reference missing',
+        output: 'Internal configuration issue. This tool needs the AI service injected at startup.',
+      };
+    }
+
+    const { query, instructions, model } = params || {};
+    if (!query) {
+      return {
+        success: false,
+        error: 'query is required',
+        output: 'Provide a `query` parameter with the search request.',
+      };
+    }
+
+    try {
+      // Idempotency key derived from agent + content hash; lets a
+      // repeated identical query inside the same turn reuse the
+      // prior billing row (the backend dedups on the same key).
+      const idempotencyKey = this._buildIdempotencyKey(context, query);
+
+      const result = await this.aiService.foundryWebSearch(query, {
+        instructions,
+        model,
+        idempotencyKey,
+        sessionId: context?.sessionId,
+        platformProvided: context?.platformProvided !== false,
+      });
+
+      // Format for agent consumption. The answer goes in `output` so
+      // the agent reads it as the primary result; structured fields
+      // (citations, usage) live alongside for any downstream tool
+      // that wants to act on URLs.
+      const citationsBlock = (result.citations || []).length > 0
+        ? '\n\nSources:\n' + result.citations.map((c, i) =>
+            `  [${i + 1}] ${c.title || '(untitled)'} — ${c.url}`).join('\n')
+        : '\n\n(No sources cited.)';
+
+      const searchCount = result.usage?.webSearches || 0;
+      const meta = ` [searches=${searchCount}, tokens=${result.usage?.inputTokens || 0}in/${result.usage?.outputTokens || 0}out]`;
+
+      return {
+        success: true,
+        output: result.answer + citationsBlock + meta,
+        // Structured fields downstream tools can inspect:
+        answer: result.answer,
+        citations: result.citations,
+        usage: result.usage,
+        model: result.model,
+        // Hint to the agent: if webSearches=0, the model didn't
+        // actually search. Surface this so the agent can decide to
+        // retry with stronger wording.
+        ...(searchCount === 0 ? { warning: 'Model did not invoke the web search tool. Try phrasing the query as "Search the web for: ..." to force tool use.' } : {}),
+      };
+
+    } catch (error) {
+      // 503 = feature flag off / not provisioned. Distinct error path
+      // so the agent can fall back to the `web` tool instead of
+      // retrying.
+      if (error.status === 503) {
+        return {
+          success: false,
+          error: error.message,
+          output: error.message + '\n\nFallback: use the `web` tool for puppeteer-based search instead.',
+          notProvisioned: true,
+        };
+      }
+      return {
+        success: false,
+        error: error.message,
+        output: `Web search failed: ${error.message}`,
+      };
+    }
+  }
+
+  /**
+   * Build an idempotency key per (agentId, query). Stable for repeated
+   * identical queries inside the same turn — the backend's
+   * Transaction.findOne dedup short-circuits the duplicate billing.
+   *
+   * Different turns or different agents → different keys → fresh
+   * search + bill, which is correct (the world moves on between
+   * turns, the user expects a fresh result).
+   *
+   * @private
+   */
+  _buildIdempotencyKey(context, query) {
+    const agentId = context?.agentId || 'unknown';
+    const turn = context?.turnId || context?.messageId || Date.now();
+    // Cheap, deterministic, no-deps hash. Doesn't need to be
+    // cryptographic — the keyspace is per-agent so collision risk
+    // is minimal.
+    let h = 0;
+    const s = `${agentId}|${turn}|${query}`;
+    for (let i = 0; i < s.length; i++) {
+      h = ((h << 5) - h) + s.charCodeAt(i);
+      h |= 0;
+    }
+    return `fws-${agentId}-${(h >>> 0).toString(36)}`;
+  }
+}
+
+export default FoundryWebSearchTool;