graphlit-client 1.0.20250612002 → 1.0.20250612003

package/dist/client.js

@@ -1566,12 +1566,7 @@ class Graphlit {
                 ? (await this.getSpecification(specification.id))
                     .specification
                 : undefined;
-            // Check streaming support
-            if (fullSpec && !this.supportsStreaming(fullSpec)) {
-                throw new Error("Streaming is not supported for this specification. " +
-                    "Use promptAgent() instead or configure a streaming client.");
-            }
-            // Ensure conversation
+            // Ensure conversation exists first (before streaming check)
             let actualConversationId = conversationId;
             if (!actualConversationId) {
                 const createResponse = await this.createConversation({
@@ -1586,6 +1581,47 @@ class Graphlit {
                     throw new Error("Failed to create conversation");
                 }
             }
+            // Check streaming support - fallback to promptAgent if not supported
+            if (fullSpec && !this.supportsStreaming(fullSpec)) {
+                if (process.env.DEBUG_GRAPHLIT_STREAMING) {
+                    console.log("\n⚠️ [streamAgent] Streaming not supported, falling back to promptAgent with same conversation");
+                }
+                // Fallback to promptAgent using the same conversation and parameters
+                const promptResult = await this.promptAgent(prompt, actualConversationId, // Preserve conversation
+                specification, tools, toolHandlers, {
+                    maxToolRounds: maxRounds,
+                }, mimeType, data, contentFilter, augmentedFilter, correlationId);
+                // Convert promptAgent result to streaming events
+                onEvent({
+                    type: "conversation_started",
+                    conversationId: actualConversationId,
+                    timestamp: new Date(),
+                });
+                // Emit the final message as a single update (simulating streaming)
+                onEvent({
+                    type: "message_update",
+                    message: {
+                        __typename: "ConversationMessage",
+                        message: promptResult.message,
+                        role: Types.ConversationRoleTypes.Assistant,
+                        timestamp: new Date().toISOString(),
+                        toolCalls: [],
+                    },
+                    isStreaming: false,
+                });
+                // Emit completion event
+                onEvent({
+                    type: "conversation_completed",
+                    message: {
+                        __typename: "ConversationMessage",
+                        message: promptResult.message,
+                        role: Types.ConversationRoleTypes.Assistant,
+                        timestamp: new Date().toISOString(),
+                        toolCalls: [],
+                    },
+                });
+                return; // Exit early after successful fallback
+            }
             // Create UI event adapter
             uiAdapter = new UIEventAdapter(onEvent, actualConversationId, {
                 smoothingEnabled: options?.smoothingEnabled ?? true,

package/dist/generated/graphql-documents.js

@@ -636,6 +636,7 @@ export const DescribeEncodedImage = gql `
         posts
         chapters
         questions
+        quotes
         video {
           width
           height
@@ -770,6 +771,7 @@ export const DescribeImage = gql `
         posts
         chapters
         questions
+        quotes
         video {
           width
           height
@@ -1625,6 +1627,7 @@ export const PublishContents = gql `
       posts
       chapters
       questions
+      quotes
       video {
         width
         height
@@ -1749,6 +1752,7 @@ export const PublishText = gql `
       posts
       chapters
       questions
+      quotes
       video {
         width
         height
@@ -1880,6 +1884,7 @@ export const QueryContents = gql `
       posts
       chapters
       questions
+      quotes
       video {
         width
         height
@@ -2316,6 +2321,7 @@ export const AskGraphlit = gql `
           posts
           chapters
           questions
+          quotes
           video {
             width
             height
@@ -2485,6 +2491,7 @@ export const CompleteConversation = gql `
           posts
           chapters
           questions
+          quotes
           video {
             width
             height
@@ -2662,6 +2669,7 @@ export const CompleteConversation = gql `
             posts
             chapters
             questions
+            quotes
             video {
               width
               height
@@ -2801,6 +2809,7 @@ export const ContinueConversation = gql `
           posts
           chapters
           questions
+          quotes
           video {
             width
             height
@@ -2978,6 +2987,7 @@ export const ContinueConversation = gql `
             posts
             chapters
             questions
+            quotes
             video {
               width
               height
@@ -3165,6 +3175,7 @@ export const FormatConversation = gql `
           posts
           chapters
           questions
+          quotes
           video {
             width
             height
@@ -3342,6 +3353,7 @@ export const FormatConversation = gql `
             posts
             chapters
             questions
+            quotes
             video {
               width
               height
@@ -3484,6 +3496,7 @@ export const GetConversation = gql `
           posts
           chapters
           questions
+          quotes
           video {
             width
             height
@@ -3799,6 +3812,7 @@ export const Prompt = gql `
           posts
           chapters
           questions
+          quotes
           video {
             width
             height
@@ -3944,6 +3958,7 @@ export const PromptConversation = gql `
           posts
           chapters
           questions
+          quotes
           video {
             width
             height
@@ -4121,6 +4136,7 @@ export const PromptConversation = gql `
             posts
             chapters
             questions
+            quotes
             video {
               width
               height
@@ -4256,6 +4272,7 @@ export const PublishConversation = gql `
       posts
       chapters
       questions
+      quotes
       video {
         width
         height
@@ -4387,6 +4404,7 @@ export const QueryConversations = gql `
             posts
             chapters
             questions
+            quotes
             video {
               width
               height
@@ -4728,6 +4746,7 @@ export const ReviseContent = gql `
           posts
           chapters
           questions
+          quotes
           video {
             width
             height
@@ -4870,6 +4889,7 @@ export const ReviseEncodedImage = gql `
           posts
           chapters
           questions
+          quotes
           video {
             width
             height
@@ -5011,6 +5031,7 @@ export const ReviseImage = gql `
           posts
           chapters
           questions
+          quotes
           video {
             width
             height
@@ -5152,6 +5173,7 @@ export const ReviseText = gql `
           posts
           chapters
           questions
+          quotes
           video {
             width
             height
@@ -8187,6 +8209,7 @@ export const PromptSpecifications = gql `
           posts
           chapters
           questions
+          quotes
           video {
             width
             height

package/dist/generated/graphql-types.d.ts

@@ -1645,6 +1645,8 @@ export type Content = {
     posts?: Maybe<Array<Scalars['String']['output']>>;
     /** The followup questions which can be asked about the content. */
     questions?: Maybe<Array<Scalars['String']['output']>>;
+    /** Quotes extracted from the content. */
+    quotes?: Maybe<Array<Scalars['String']['output']>>;
     /** The relevance score of the content. */
     relevance?: Maybe<Scalars['Float']['output']>;
     /** The renditions generated from this content. */
@@ -2206,6 +2208,8 @@ export type ContentUpdateInput = {
     posts?: InputMaybe<Array<Scalars['String']['input']>>;
     /** The followup questions which can be asked about the content. */
     questions?: InputMaybe<Array<Scalars['String']['input']>>;
+    /** Quotes extracted from the content. */
+    quotes?: InputMaybe<Array<Scalars['String']['input']>>;
     /** The content shape metadata. */
     shape?: InputMaybe<ShapeMetadataInput>;
     /** The content summary. */
@@ -2529,6 +2533,8 @@ export type ConversationUpdateInput = {
     filter?: InputMaybe<ContentCriteriaInput>;
     /** The ID of the conversation to update. */
     id: Scalars['ID']['input'];
+    /** The conversation messages. */
+    messages?: InputMaybe<Array<ConversationMessageInput>>;
     /** The name of the conversation. */
     name?: InputMaybe<Scalars['String']['input']>;
     /** The LLM specification used by this conversation, optional. */
@@ -13039,6 +13045,8 @@ export declare enum SummarizationTypes {
     Posts = "POSTS",
     /** Questions */
     Questions = "QUESTIONS",
+    /** Quote */
+    Quotes = "QUOTES",
     /** Summary */
     Summary = "SUMMARY"
 }
@@ -14739,6 +14747,7 @@ export type DescribeEncodedImageMutation = {
                 posts?: Array<string> | null;
                 chapters?: Array<string> | null;
                 questions?: Array<string> | null;
+                quotes?: Array<string> | null;
                 video?: {
                     __typename?: 'VideoMetadata';
                     width?: number | null;
@@ -14881,6 +14890,7 @@ export type DescribeImageMutation = {
                 posts?: Array<string> | null;
                 chapters?: Array<string> | null;
                 questions?: Array<string> | null;
+                quotes?: Array<string> | null;
                 video?: {
                     __typename?: 'VideoMetadata';
                     width?: number | null;
@@ -15821,6 +15831,7 @@ export type PublishContentsMutation = {
             posts?: Array<string> | null;
             chapters?: Array<string> | null;
             questions?: Array<string> | null;
+            quotes?: Array<string> | null;
             video?: {
                 __typename?: 'VideoMetadata';
                 width?: number | null;
@@ -15953,6 +15964,7 @@ export type PublishTextMutation = {
             posts?: Array<string> | null;
             chapters?: Array<string> | null;
             questions?: Array<string> | null;
+            quotes?: Array<string> | null;
             video?: {
                 __typename?: 'VideoMetadata';
                 width?: number | null;
@@ -16081,6 +16093,7 @@ export type QueryContentsQuery = {
             posts?: Array<string> | null;
             chapters?: Array<string> | null;
             questions?: Array<string> | null;
+            quotes?: Array<string> | null;
             error?: string | null;
             owner: {
                 __typename?: 'Owner';
@@ -16623,6 +16636,7 @@ export type AskGraphlitMutation = {
                     posts?: Array<string> | null;
                     chapters?: Array<string> | null;
                     questions?: Array<string> | null;
+                    quotes?: Array<string> | null;
                     video?: {
                         __typename?: 'VideoMetadata';
                         width?: number | null;
@@ -16811,6 +16825,7 @@ export type CompleteConversationMutation = {
                     posts?: Array<string> | null;
                     chapters?: Array<string> | null;
                     questions?: Array<string> | null;
+                    quotes?: Array<string> | null;
                     video?: {
                         __typename?: 'VideoMetadata';
                         width?: number | null;
@@ -17003,6 +17018,7 @@ export type CompleteConversationMutation = {
                         posts?: Array<string> | null;
                         chapters?: Array<string> | null;
                         questions?: Array<string> | null;
+                        quotes?: Array<string> | null;
                         video?: {
                             __typename?: 'VideoMetadata';
                             width?: number | null;
@@ -17153,6 +17169,7 @@ export type ContinueConversationMutation = {
                     posts?: Array<string> | null;
                     chapters?: Array<string> | null;
                     questions?: Array<string> | null;
+                    quotes?: Array<string> | null;
                     video?: {
                         __typename?: 'VideoMetadata';
                         width?: number | null;
@@ -17345,6 +17362,7 @@ export type ContinueConversationMutation = {
                         posts?: Array<string> | null;
                         chapters?: Array<string> | null;
                         questions?: Array<string> | null;
+                        quotes?: Array<string> | null;
                         video?: {
                             __typename?: 'VideoMetadata';
                             width?: number | null;
@@ -17559,6 +17577,7 @@ export type FormatConversationMutation = {
                     posts?: Array<string> | null;
                     chapters?: Array<string> | null;
                     questions?: Array<string> | null;
+                    quotes?: Array<string> | null;
                     video?: {
                         __typename?: 'VideoMetadata';
                         width?: number | null;
@@ -17751,6 +17770,7 @@ export type FormatConversationMutation = {
                         posts?: Array<string> | null;
                         chapters?: Array<string> | null;
                         questions?: Array<string> | null;
+                        quotes?: Array<string> | null;
                         video?: {
                             __typename?: 'VideoMetadata';
                             width?: number | null;
@@ -17906,6 +17926,7 @@ export type GetConversationQuery = {
                     posts?: Array<string> | null;
                     chapters?: Array<string> | null;
                     questions?: Array<string> | null;
+                    quotes?: Array<string> | null;
                     video?: {
                         __typename?: 'VideoMetadata';
                         width?: number | null;
@@ -18284,6 +18305,7 @@ export type PromptMutation = {
                     posts?: Array<string> | null;
                     chapters?: Array<string> | null;
                     questions?: Array<string> | null;
+                    quotes?: Array<string> | null;
                     video?: {
                         __typename?: 'VideoMetadata';
                         width?: number | null;
@@ -18439,6 +18461,7 @@ export type PromptConversationMutation = {
                     posts?: Array<string> | null;
                     chapters?: Array<string> | null;
                     questions?: Array<string> | null;
+                    quotes?: Array<string> | null;
                     video?: {
                         __typename?: 'VideoMetadata';
                         width?: number | null;
@@ -18631,6 +18654,7 @@ export type PromptConversationMutation = {
                         posts?: Array<string> | null;
                         chapters?: Array<string> | null;
                         questions?: Array<string> | null;
+                        quotes?: Array<string> | null;
                         video?: {
                             __typename?: 'VideoMetadata';
                             width?: number | null;
@@ -18757,6 +18781,7 @@ export type PublishConversationMutation = {
             posts?: Array<string> | null;
             chapters?: Array<string> | null;
             questions?: Array<string> | null;
+            quotes?: Array<string> | null;
             video?: {
                 __typename?: 'VideoMetadata';
                 width?: number | null;
@@ -18919,6 +18944,7 @@ export type QueryConversationsQuery = {
                         posts?: Array<string> | null;
                         chapters?: Array<string> | null;
                         questions?: Array<string> | null;
+                        quotes?: Array<string> | null;
                         video?: {
                             __typename?: 'VideoMetadata';
                             width?: number | null;
@@ -19326,6 +19352,7 @@ export type ReviseContentMutation = {
                     posts?: Array<string> | null;
                     chapters?: Array<string> | null;
                     questions?: Array<string> | null;
+                    quotes?: Array<string> | null;
                     video?: {
                         __typename?: 'VideoMetadata';
                         width?: number | null;
@@ -19478,6 +19505,7 @@ export type ReviseEncodedImageMutation = {
                     posts?: Array<string> | null;
                     chapters?: Array<string> | null;
                     questions?: Array<string> | null;
+                    quotes?: Array<string> | null;
                     video?: {
                         __typename?: 'VideoMetadata';
                         width?: number | null;
@@ -19629,6 +19657,7 @@ export type ReviseImageMutation = {
                     posts?: Array<string> | null;
                     chapters?: Array<string> | null;
                     questions?: Array<string> | null;
+                    quotes?: Array<string> | null;
                     video?: {
                         __typename?: 'VideoMetadata';
                         width?: number | null;
@@ -19780,6 +19809,7 @@ export type ReviseTextMutation = {
                     posts?: Array<string> | null;
                     chapters?: Array<string> | null;
                     questions?: Array<string> | null;
+                    quotes?: Array<string> | null;
                     video?: {
                         __typename?: 'VideoMetadata';
                         width?: number | null;
@@ -23592,6 +23622,7 @@ export type PromptSpecificationsMutation = {
                     posts?: Array<string> | null;
                     chapters?: Array<string> | null;
                     questions?: Array<string> | null;
+                    quotes?: Array<string> | null;
                     video?: {
                         __typename?: 'VideoMetadata';
                         width?: number | null;

package/dist/generated/graphql-types.js

@@ -1866,6 +1866,8 @@ export var SummarizationTypes;
     SummarizationTypes["Posts"] = "POSTS";
     /** Questions */
     SummarizationTypes["Questions"] = "QUESTIONS";
+    /** Quote */
+    SummarizationTypes["Quotes"] = "QUOTES";
     /** Summary */
     SummarizationTypes["Summary"] = "SUMMARY";
 })(SummarizationTypes || (SummarizationTypes = {}));

package/dist/streaming/providers.js

@@ -21,9 +21,22 @@ onEvent, onComplete) {
     // Performance metrics
     const startTime = Date.now();
     let firstTokenTime = 0;
+    let firstMeaningfulContentTime = 0;
     let tokenCount = 0;
+    let toolArgumentTokens = 0;
     let lastEventTime = 0;
     const interTokenDelays = [];
+    // Tool calling metrics
+    const toolMetrics = {
+        totalTools: 0,
+        successfulTools: 0,
+        failedTools: 0,
+        toolTimes: [],
+        currentToolStart: 0,
+        roundStartTime: startTime,
+        rounds: [],
+        currentRound: 1
+    };
     try {
         const modelName = getModelName(specification);
         if (!modelName) {
@@ -84,13 +97,20 @@ onEvent, onComplete) {
                 fullMessage += delta.content;
                 tokenCount++;
                 const currentTime = Date.now();
-                // Track TTFT
+                // Track TTFT (first token regardless of type)
                 if (firstTokenTime === 0) {
                     firstTokenTime = currentTime - startTime;
                     if (process.env.DEBUG_GRAPHLIT_STREAMING) {
                         console.log(`\n⚡ [OpenAI] Time to First Token (TTFT): ${firstTokenTime}ms`);
                     }
                 }
+                // Track first meaningful content (excludes tool calls)
+                if (firstMeaningfulContentTime === 0 && delta.content.trim()) {
+                    firstMeaningfulContentTime = currentTime - startTime;
+                    if (process.env.DEBUG_GRAPHLIT_STREAMING) {
+                        console.log(`\n🎯 [OpenAI] Time to First Meaningful Content: ${firstMeaningfulContentTime}ms`);
+                    }
+                }
                 // Track inter-token delays
                 if (lastEventTime > 0) {
                     const delay = currentTime - lastEventTime;
@@ -115,6 +135,22 @@ onEvent, onComplete) {
                             name: "",
                             arguments: "",
                         };
+                        // Track tool metrics
+                        toolMetrics.totalTools++;
+                        toolMetrics.currentToolStart = Date.now();
+                        toolMetrics.toolTimes.push({
+                            name: toolCallDelta.function?.name || "unknown",
+                            startTime: toolMetrics.currentToolStart,
+                            argumentBuildTime: 0,
+                            totalTime: 0
+                        });
+                        // Track TTFT for first tool if no content yet
+                        if (firstTokenTime === 0) {
+                            firstTokenTime = Date.now() - startTime;
+                            if (process.env.DEBUG_GRAPHLIT_STREAMING) {
+                                console.log(`\n⚡ [OpenAI] Time to First Token (Tool Call): ${firstTokenTime}ms`);
+                            }
+                        }
                         if (process.env.DEBUG_GRAPHLIT_STREAMING) {
                             console.log(`[OpenAI] Starting new tool call: ${toolCalls[index].id}`);
                         }
@@ -134,6 +170,8 @@ onEvent, onComplete) {
                     }
                     if (toolCallDelta.function?.arguments) {
                         toolCalls[index].arguments += toolCallDelta.function.arguments;
+                        // Count tool argument tokens (rough estimate: ~4 chars per token)
+                        toolArgumentTokens += Math.ceil(toolCallDelta.function.arguments.length / 4);
                         // Debug logging for partial JSON accumulation
                         if (process.env.DEBUG_GRAPHLIT_STREAMING) {
                             console.log(`[OpenAI] Tool ${toolCalls[index].name} - Partial JSON chunk: "${toolCallDelta.function.arguments}"`);
@@ -148,20 +186,33 @@ onEvent, onComplete) {
                 }
             }
         }
-        // Emit complete events for tool calls
-        for (const toolCall of toolCalls) {
+        // Emit complete events for tool calls and finalize metrics
+        for (let i = 0; i < toolCalls.length; i++) {
+            const toolCall = toolCalls[i];
+            const currentTime = Date.now();
+            // Update tool metrics
+            if (i < toolMetrics.toolTimes.length) {
+                const toolTime = toolMetrics.toolTimes[i];
+                toolTime.argumentBuildTime = currentTime - toolTime.startTime;
+                toolTime.totalTime = toolTime.argumentBuildTime; // For streaming, this is the same
+                toolTime.name = toolCall.name; // Update with final name
+            }
+            // Track tool success/failure
+            try {
+                JSON.parse(toolCall.arguments);
+                toolMetrics.successfulTools++;
+                if (process.env.DEBUG_GRAPHLIT_STREAMING) {
+                    console.log(`[OpenAI] ✅ Valid JSON for ${toolCall.name}`);
+                }
+            }
+            catch (e) {
+                toolMetrics.failedTools++;
+                console.error(`[OpenAI] ❌ Invalid JSON for ${toolCall.name}: ${e}`);
+            }
             // Log the final JSON for debugging
             if (process.env.DEBUG_GRAPHLIT_STREAMING) {
                 console.log(`[OpenAI] Tool ${toolCall.name} complete with arguments (${toolCall.arguments.length} chars):`);
                 console.log(toolCall.arguments);
-                // Validate JSON
-                try {
-                    JSON.parse(toolCall.arguments);
-                    console.log(`[OpenAI] ✅ Valid JSON for ${toolCall.name}`);
-                }
-                catch (e) {
-                    console.error(`[OpenAI] ❌ Invalid JSON for ${toolCall.name}: ${e}`);
-                }
             }
             onEvent({
                 type: "tool_call_complete",
@@ -176,22 +227,60 @@ onEvent, onComplete) {
         if (process.env.DEBUG_GRAPHLIT_STREAMING && toolCalls.length > 0) {
             console.log(`[OpenAI] Successfully processed ${toolCalls.length} tool calls`);
         }
-        // Calculate final metrics
+        // Calculate final metrics including tool calling insights
         const totalTime = Date.now() - startTime;
-        const tokensPerSecond = tokenCount > 0 ? tokenCount / (totalTime / 1000) : 0;
+        const totalTokens = tokenCount + toolArgumentTokens;
+        const tokensPerSecond = totalTokens > 0 ? totalTokens / (totalTime / 1000) : 0;
+        // Finalize round metrics
+        if (toolCalls.length > 0) {
+            const roundEndTime = Date.now();
+            const totalToolTime = toolMetrics.toolTimes.reduce((sum, tool) => sum + tool.totalTime, 0);
+            const llmTime = totalTime - totalToolTime;
+            toolMetrics.rounds.push({
+                roundNumber: toolMetrics.currentRound,
+                llmTime: llmTime,
+                toolTime: totalToolTime,
+                toolCount: toolCalls.length
+            });
+        }
         if (process.env.DEBUG_GRAPHLIT_STREAMING) {
             console.log("\n📊 [OpenAI] Performance Metrics:");
             console.log(`  ⏱️  Total Time: ${totalTime}ms`);
             console.log(`  ⚡ Time to First Token (TTFT): ${firstTokenTime}ms`);
-            console.log(`  📈 Tokens Generated: ${tokenCount}`);
+            if (firstMeaningfulContentTime > 0) {
+                console.log(`  🎯 Time to First Meaningful Content: ${firstMeaningfulContentTime}ms`);
+            }
+            console.log(`  📈 Content Tokens: ${tokenCount}`);
+            console.log(`  🔧 Tool Argument Tokens: ${toolArgumentTokens}`);
+            console.log(`  📊 Total Tokens: ${totalTokens}`);
             console.log(`  💨 Tokens Per Second (TPS): ${tokensPerSecond.toFixed(2)}`);
+            // Tool calling metrics
+            if (toolCalls.length > 0) {
+                console.log(`\n🔧 [OpenAI] Tool Calling Metrics:`);
+                console.log(`  🛠️  Total Tools Called: ${toolMetrics.totalTools}`);
+                console.log(`  ✅ Successful Tools: ${toolMetrics.successfulTools}`);
+                console.log(`  ❌ Failed Tools: ${toolMetrics.failedTools}`);
+                console.log(`  📊 Success Rate: ${((toolMetrics.successfulTools / toolMetrics.totalTools) * 100).toFixed(1)}%`);
+                // Tool timing details
+                toolMetrics.toolTimes.forEach((tool, idx) => {
+                    console.log(`  🔨 Tool ${idx + 1} (${tool.name}): ${tool.argumentBuildTime}ms`);
+                });
+                const avgToolTime = toolMetrics.toolTimes.reduce((sum, tool) => sum + tool.totalTime, 0) / toolMetrics.toolTimes.length;
+                console.log(`  ⏱️  Average Tool Time: ${avgToolTime.toFixed(2)}ms`);
+                // Round metrics
+                toolMetrics.rounds.forEach(round => {
+                    const efficiency = round.toolCount > 0 ? (round.llmTime / (round.llmTime + round.toolTime) * 100).toFixed(1) : 100;
+                    console.log(`  🔄 Round ${round.roundNumber}: LLM=${round.llmTime}ms, Tools=${round.toolTime}ms (${round.toolCount} tools), Efficiency=${efficiency}%`);
+                });
+            }
             if (interTokenDelays.length > 0) {
                 const avgDelay = interTokenDelays.reduce((a, b) => a + b, 0) / interTokenDelays.length;
                 const sortedDelays = [...interTokenDelays].sort((a, b) => a - b);
                 const p50Delay = sortedDelays[Math.floor(sortedDelays.length * 0.5)];
                 const p95Delay = sortedDelays[Math.floor(sortedDelays.length * 0.95)];
                 const p99Delay = sortedDelays[Math.floor(sortedDelays.length * 0.99)];
-                console.log(`  ⏳ Average Inter-Token Delay: ${avgDelay.toFixed(2)}ms`);
+                console.log(`\n⏳ [OpenAI] Inter-Token Timing:`);
+                console.log(`  📊 Average Delay: ${avgDelay.toFixed(2)}ms`);
                 console.log(`  📊 P50 Delay: ${p50Delay}ms`);
                 console.log(`  ⚠️  P95 Delay: ${p95Delay}ms`);
                 console.log(`  🚨 P99 Delay: ${p99Delay}ms`);
@@ -215,9 +304,22 @@ onEvent, onComplete) {
     // Performance metrics
     const startTime = Date.now();
     let firstTokenTime = 0;
+    let firstMeaningfulContentTime = 0;
     let tokenCount = 0;
+    let toolArgumentTokens = 0;
     let lastEventTime = 0;
     const interTokenDelays = [];
+    // Tool calling metrics
+    const toolMetrics = {
+        totalTools: 0,
+        successfulTools: 0,
+        failedTools: 0,
+        toolTimes: [],
+        currentToolStart: 0,
+        roundStartTime: startTime,
+        rounds: [],
+        currentRound: 1
+    };
     try {
         const modelName = getModelName(specification);
         if (!modelName) {
@@ -271,6 +373,22 @@ onEvent, onComplete) {
                         arguments: "",
                     };
                     toolCalls.push(toolCall);
+                    // Track tool metrics
+                    toolMetrics.totalTools++;
+                    toolMetrics.currentToolStart = Date.now();
+                    toolMetrics.toolTimes.push({
+                        name: toolCall.name,
+                        startTime: toolMetrics.currentToolStart,
+                        argumentBuildTime: 0,
+                        totalTime: 0
+                    });
+                    // Track TTFT for first tool if no content yet
+                    if (firstTokenTime === 0) {
+                        firstTokenTime = Date.now() - startTime;
+                        if (process.env.DEBUG_GRAPHLIT_STREAMING) {
+                            console.log(`\n⚡ [Anthropic] Time to First Token (Tool Call): ${firstTokenTime}ms`);
+                        }
+                    }
                     onEvent({
                         type: "tool_call_start",
                         toolCall: {
@@ -285,13 +403,20 @@ onEvent, onComplete) {
                     fullMessage += chunk.delta.text;
                     tokenCount++;
                     const currentTime = Date.now();
-                    // Track TTFT
+                    // Track TTFT (first token regardless of type)
                     if (firstTokenTime === 0) {
                         firstTokenTime = currentTime - startTime;
                         if (process.env.DEBUG_GRAPHLIT_STREAMING) {
                             console.log(`\n⚡ [Anthropic] Time to First Token (TTFT): ${firstTokenTime}ms`);
                         }
                     }
+                    // Track first meaningful content (excludes tool calls)
+                    if (firstMeaningfulContentTime === 0 && chunk.delta.text.trim()) {
+                        firstMeaningfulContentTime = currentTime - startTime;
+                        if (process.env.DEBUG_GRAPHLIT_STREAMING) {
+                            console.log(`\n🎯 [Anthropic] Time to First Meaningful Content: ${firstMeaningfulContentTime}ms`);
+                        }
+                    }
                     // Track inter-token delays
                     if (lastEventTime > 0) {
                         const delay = currentTime - lastEventTime;
@@ -311,6 +436,8 @@ onEvent, onComplete) {
                     const currentTool = toolCalls[toolCalls.length - 1];
                     if (currentTool) {
                         currentTool.arguments += chunk.delta.partial_json;
+                        // Count tool argument tokens (rough estimate: ~4 chars per token)
+                        toolArgumentTokens += Math.ceil(chunk.delta.partial_json.length / 4);
                         // Debug logging for partial JSON accumulation
                         if (process.env.DEBUG_GRAPHLIT_STREAMING) {
                             console.log(`[Anthropic] Tool ${currentTool.name} - Partial JSON chunk: "${chunk.delta.partial_json}"`);
@@ -329,6 +456,27 @@ onEvent, onComplete) {
                 // Tool call complete
                 const currentTool = toolCalls[toolCalls.length - 1];
                 if (currentTool) {
+                    const currentTime = Date.now();
+                    // Update tool metrics
+                    const toolIndex = toolCalls.length - 1;
+                    if (toolIndex < toolMetrics.toolTimes.length) {
+                        const toolTime = toolMetrics.toolTimes[toolIndex];
+                        toolTime.argumentBuildTime = currentTime - toolTime.startTime;
+                        toolTime.totalTime = toolTime.argumentBuildTime;
+                        toolTime.name = currentTool.name;
+                    }
+                    // Track tool success/failure
+                    try {
+                        JSON.parse(currentTool.arguments);
+                        toolMetrics.successfulTools++;
+                        if (process.env.DEBUG_GRAPHLIT_STREAMING) {
+                            console.log(`[Anthropic] ✅ Valid JSON for ${currentTool.name}`);
+                        }
+                    }
+                    catch (e) {
+                        toolMetrics.failedTools++;
+                        console.error(`[Anthropic] ❌ Invalid JSON for ${currentTool.name}: ${e}`);
+                    }
                     // Log the final JSON for debugging
                     if (process.env.DEBUG_GRAPHLIT_STREAMING ||
                         !isValidJSON(currentTool.arguments)) {
@@ -340,16 +488,6 @@ onEvent, onComplete) {
                             currentTool.arguments.length > 100) {
                             console.warn(`[Anthropic] WARNING: JSON may be truncated - doesn't end with '}': ...${lastChars}`);
                         }
-                        // Validate JSON
-                        try {
-                            JSON.parse(currentTool.arguments);
-                            if (process.env.DEBUG_GRAPHLIT_STREAMING) {
-                                console.log(`[Anthropic] ✅ Valid JSON for ${currentTool.name}`);
-                            }
-                        }
-                        catch (e) {
-                            console.error(`[Anthropic] ❌ Invalid JSON for ${currentTool.name}: ${e}`);
-                        }
                     }
                     onEvent({
                         type: "tool_call_complete",
@@ -399,22 +537,60 @@ onEvent, onComplete) {
             console.log(`[Anthropic] Filtered out ${toolCalls.length - validToolCalls.length} incomplete tool calls`);
             console.log(`[Anthropic] Successfully processed ${validToolCalls.length} valid tool calls`);
         }
-        // Calculate final metrics
+        // Calculate final metrics including tool calling insights
         const totalTime = Date.now() - startTime;
-        const tokensPerSecond = tokenCount > 0 ? tokenCount / (totalTime / 1000) : 0;
+        const totalTokens = tokenCount + toolArgumentTokens;
+        const tokensPerSecond = totalTokens > 0 ? totalTokens / (totalTime / 1000) : 0;
+        // Finalize round metrics
+        if (validToolCalls.length > 0) {
+            const roundEndTime = Date.now();
+            const totalToolTime = toolMetrics.toolTimes.reduce((sum, tool) => sum + tool.totalTime, 0);
+            const llmTime = totalTime - totalToolTime;
+            toolMetrics.rounds.push({
+                roundNumber: toolMetrics.currentRound,
+                llmTime: llmTime,
+                toolTime: totalToolTime,
+                toolCount: validToolCalls.length
+            });
+        }
         if (process.env.DEBUG_GRAPHLIT_STREAMING) {
             console.log("\n📊 [Anthropic] Performance Metrics:");
             console.log(`  ⏱️  Total Time: ${totalTime}ms`);
             console.log(`  ⚡ Time to First Token (TTFT): ${firstTokenTime}ms`);
-            console.log(`  📈 Tokens Generated: ${tokenCount}`);
+            if (firstMeaningfulContentTime > 0) {
+                console.log(`  🎯 Time to First Meaningful Content: ${firstMeaningfulContentTime}ms`);
+            }
+            console.log(`  📈 Content Tokens: ${tokenCount}`);
+            console.log(`  🔧 Tool Argument Tokens: ${toolArgumentTokens}`);
+            console.log(`  📊 Total Tokens: ${totalTokens}`);
             console.log(`  💨 Tokens Per Second (TPS): ${tokensPerSecond.toFixed(2)}`);
+            // Tool calling metrics
+            if (validToolCalls.length > 0) {
+                console.log(`\n🔧 [Anthropic] Tool Calling Metrics:`);
+                console.log(`  🛠️  Total Tools Called: ${toolMetrics.totalTools}`);
+                console.log(`  ✅ Successful Tools: ${toolMetrics.successfulTools}`);
+                console.log(`  ❌ Failed Tools: ${toolMetrics.failedTools}`);
+                console.log(`  📊 Success Rate: ${((toolMetrics.successfulTools / toolMetrics.totalTools) * 100).toFixed(1)}%`);
+                // Tool timing details
+                toolMetrics.toolTimes.forEach((tool, idx) => {
+                    console.log(`  🔨 Tool ${idx + 1} (${tool.name}): ${tool.argumentBuildTime}ms`);
+                });
+                const avgToolTime = toolMetrics.toolTimes.reduce((sum, tool) => sum + tool.totalTime, 0) / toolMetrics.toolTimes.length;
+                console.log(`  ⏱️  Average Tool Time: ${avgToolTime.toFixed(2)}ms`);
+                // Round metrics
+                toolMetrics.rounds.forEach(round => {
+                    const efficiency = round.toolCount > 0 ? (round.llmTime / (round.llmTime + round.toolTime) * 100).toFixed(1) : 100;
+                    console.log(`  🔄 Round ${round.roundNumber}: LLM=${round.llmTime}ms, Tools=${round.toolTime}ms (${round.toolCount} tools), Efficiency=${efficiency}%`);
+                });
+            }
             if (interTokenDelays.length > 0) {
                 const avgDelay = interTokenDelays.reduce((a, b) => a + b, 0) / interTokenDelays.length;
                 const sortedDelays = [...interTokenDelays].sort((a, b) => a - b);
                 const p50Delay = sortedDelays[Math.floor(sortedDelays.length * 0.5)];
                 const p95Delay = sortedDelays[Math.floor(sortedDelays.length * 0.95)];
                 const p99Delay = sortedDelays[Math.floor(sortedDelays.length * 0.99)];
-                console.log(`  ⏳ Average Inter-Token Delay: ${avgDelay.toFixed(2)}ms`);
+                console.log(`\n⏳ [Anthropic] Inter-Token Timing:`);
+                console.log(`  📊 Average Delay: ${avgDelay.toFixed(2)}ms`);
                 console.log(`  📊 P50 Delay: ${p50Delay}ms`);
                 console.log(`  ⚠️  P95 Delay: ${p95Delay}ms`);
                 console.log(`  🚨 P99 Delay: ${p99Delay}ms`);
@@ -438,9 +614,22 @@ onEvent, onComplete) {
     // Performance metrics
     const startTime = Date.now();
     let firstTokenTime = 0;
+    let firstMeaningfulContentTime = 0;
     let tokenCount = 0;
+    let toolArgumentTokens = 0;
     let lastEventTime = 0;
     const interTokenDelays = [];
+    // Tool calling metrics
+    const toolMetrics = {
+        totalTools: 0,
+        successfulTools: 0,
+        failedTools: 0,
+        toolTimes: [],
+        currentToolStart: 0,
+        roundStartTime: startTime,
+        rounds: [],
+        currentRound: 1
+    };
     try {
         const modelName = getModelName(specification);
         if (!modelName) {
@@ -514,6 +703,22 @@ onEvent, onComplete) {
             }
             if (text) {
                 fullMessage += text;
+                tokenCount++;
+                const currentTime = Date.now();
+                // Track TTFT (first token regardless of type)
+                if (firstTokenTime === 0) {
+                    firstTokenTime = currentTime - startTime;
+                    if (process.env.DEBUG_GRAPHLIT_STREAMING) {
+                        console.log(`\n⚡ [Google] Time to First Token (TTFT): ${firstTokenTime}ms`);
+                    }
+                }
+                // Track first meaningful content
+                if (firstMeaningfulContentTime === 0 && text.trim()) {
+                    firstMeaningfulContentTime = currentTime - startTime;
+                    if (process.env.DEBUG_GRAPHLIT_STREAMING) {
+                        console.log(`\n🎯 [Google] Time to First Meaningful Content: ${firstMeaningfulContentTime}ms`);
+                    }
+                }
                 onEvent({
                     type: "token",
                     token: text,
@@ -536,6 +741,23 @@ onEvent, onComplete) {
                                 arguments: JSON.stringify(part.functionCall.args || {}),
                             };
                             toolCalls.push(toolCall);
+                            // Track tool metrics
+                            toolMetrics.totalTools++;
+                            const argumentString = JSON.stringify(part.functionCall.args || {});
+                            toolArgumentTokens += Math.ceil(argumentString.length / 4);
+                            toolMetrics.toolTimes.push({
+                                name: part.functionCall.name,
+                                startTime: Date.now(),
+                                argumentBuildTime: 0, // Google returns complete args at once
+                                totalTime: 0
+                            });
+                            // Track TTFT for first tool if no content yet
+                            if (firstTokenTime === 0) {
+                                firstTokenTime = Date.now() - startTime;
+                                if (process.env.DEBUG_GRAPHLIT_STREAMING) {
+                                    console.log(`\n⚡ [Google] Time to First Token (Tool Call): ${firstTokenTime}ms`);
+                                }
+                            }
                             // Emit tool call events
                             onEvent({
                                 type: "tool_call_start",
@@ -549,18 +771,28 @@ onEvent, onComplete) {
                                 toolCallId: toolCall.id,
                                 argumentDelta: toolCall.arguments,
                             });
-                            // Log completion and validate JSON
+                            // Update tool metrics and validate JSON
+                            const toolIndex = toolCalls.length - 1;
+                            if (toolIndex < toolMetrics.toolTimes.length) {
+                                const toolTime = toolMetrics.toolTimes[toolIndex];
+                                toolTime.totalTime = Date.now() - toolTime.startTime;
+                                toolTime.argumentBuildTime = toolTime.totalTime; // Google returns complete args
+                            }
+                            try {
+                                JSON.parse(toolCall.arguments);
+                                toolMetrics.successfulTools++;
+                                if (process.env.DEBUG_GRAPHLIT_STREAMING) {
+                                    console.log(`[Google] ✅ Valid JSON for ${toolCall.name}`);
+                                }
+                            }
+                            catch (e) {
+                                toolMetrics.failedTools++;
+                                console.error(`[Google] ❌ Invalid JSON for ${toolCall.name}: ${e}`);
+                            }
+                            // Log completion
                             if (process.env.DEBUG_GRAPHLIT_STREAMING) {
                                 console.log(`[Google] Tool ${toolCall.name} complete with arguments (${toolCall.arguments.length} chars):`);
                                 console.log(toolCall.arguments);
-                                // Validate JSON
-                                try {
-                                    JSON.parse(toolCall.arguments);
-                                    console.log(`[Google] ✅ Valid JSON for ${toolCall.name}`);
-                                }
-                                catch (e) {
-                                    console.error(`[Google] ❌ Invalid JSON for ${toolCall.name}: ${e}`);
-                                }
                             }
                             onEvent({
                                 type: "tool_call_complete",
@@ -647,22 +879,60 @@ onEvent, onComplete) {
         if (process.env.DEBUG_GRAPHLIT_STREAMING && toolCalls.length > 0) {
             console.log(`[Google] Successfully processed ${toolCalls.length} tool calls`);
         }
-        // Calculate final metrics
+        // Calculate final metrics including tool calling insights
         const totalTime = Date.now() - startTime;
-        const tokensPerSecond = tokenCount > 0 ? tokenCount / (totalTime / 1000) : 0;
+        const totalTokens = tokenCount + toolArgumentTokens;
+        const tokensPerSecond = totalTokens > 0 ? totalTokens / (totalTime / 1000) : 0;
+        // Finalize round metrics
+        if (toolCalls.length > 0) {
+            const roundEndTime = Date.now();
+            const totalToolTime = toolMetrics.toolTimes.reduce((sum, tool) => sum + tool.totalTime, 0);
+            const llmTime = totalTime - totalToolTime;
+            toolMetrics.rounds.push({
+                roundNumber: toolMetrics.currentRound,
+                llmTime: llmTime,
+                toolTime: totalToolTime,
+                toolCount: toolCalls.length
+            });
+        }
         if (process.env.DEBUG_GRAPHLIT_STREAMING) {
             console.log("\n📊 [Google] Performance Metrics:");
             console.log(`  ⏱️  Total Time: ${totalTime}ms`);
             console.log(`  ⚡ Time to First Token (TTFT): ${firstTokenTime}ms`);
-            console.log(`  📈 Tokens Generated: ${tokenCount}`);
+            if (firstMeaningfulContentTime > 0) {
+                console.log(`  🎯 Time to First Meaningful Content: ${firstMeaningfulContentTime}ms`);
+            }
+            console.log(`  📈 Content Tokens: ${tokenCount}`);
+            console.log(`  🔧 Tool Argument Tokens: ${toolArgumentTokens}`);
+            console.log(`  📊 Total Tokens: ${totalTokens}`);
             console.log(`  💨 Tokens Per Second (TPS): ${tokensPerSecond.toFixed(2)}`);
+            // Tool calling metrics
+            if (toolCalls.length > 0) {
+                console.log(`\n🔧 [Google] Tool Calling Metrics:`);
+                console.log(`  🛠️  Total Tools Called: ${toolMetrics.totalTools}`);
+                console.log(`  ✅ Successful Tools: ${toolMetrics.successfulTools}`);
+                console.log(`  ❌ Failed Tools: ${toolMetrics.failedTools}`);
+                console.log(`  📊 Success Rate: ${((toolMetrics.successfulTools / toolMetrics.totalTools) * 100).toFixed(1)}%`);
+                // Tool timing details
+                toolMetrics.toolTimes.forEach((tool, idx) => {
+                    console.log(`  🔨 Tool ${idx + 1} (${tool.name}): ${tool.argumentBuildTime}ms`);
+                });
+                const avgToolTime = toolMetrics.toolTimes.reduce((sum, tool) => sum + tool.totalTime, 0) / toolMetrics.toolTimes.length;
+                console.log(`  ⏱️  Average Tool Time: ${avgToolTime.toFixed(2)}ms`);
+                // Round metrics
+                toolMetrics.rounds.forEach(round => {
+                    const efficiency = round.toolCount > 0 ? (round.llmTime / (round.llmTime + round.toolTime) * 100).toFixed(1) : 100;
+                    console.log(`  🔄 Round ${round.roundNumber}: LLM=${round.llmTime}ms, Tools=${round.toolTime}ms (${round.toolCount} tools), Efficiency=${efficiency}%`);
+                });
+            }
             if (interTokenDelays.length > 0) {
                 const avgDelay = interTokenDelays.reduce((a, b) => a + b, 0) / interTokenDelays.length;
                 const sortedDelays = [...interTokenDelays].sort((a, b) => a - b);
                 const p50Delay = sortedDelays[Math.floor(sortedDelays.length * 0.5)];
                 const p95Delay = sortedDelays[Math.floor(sortedDelays.length * 0.95)];
                 const p99Delay = sortedDelays[Math.floor(sortedDelays.length * 0.99)];
-                console.log(`  ⏳ Average Inter-Token Delay: ${avgDelay.toFixed(2)}ms`);
+                console.log(`\n⏳ [Google] Inter-Token Timing:`);
+                console.log(`  📊 Average Delay: ${avgDelay.toFixed(2)}ms`);
                 console.log(`  📊 P50 Delay: ${p50Delay}ms`);
                 console.log(`  ⚠️  P95 Delay: ${p95Delay}ms`);
                 console.log(`  🚨 P99 Delay: ${p99Delay}ms`);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "graphlit-client",
-  "version": "1.0.20250612002",
+  "version": "1.0.20250612003",
   "description": "Graphlit API Client for TypeScript",
   "type": "module",
   "main": "./dist/client.js",