session summarization

packages/orchestration-service/src/services/summarization.js

@@ -9,30 +9,34 @@ const MAX_SUMMARY_TOKENS = parseInt(getEnv('SUMMARY_MAX_TOKENS', SUMMARIES.MAX_S
 const MIN_EPISODES_SINCE = parseInt(getEnv('SUMMARY_MIN_EPISODES', SUMMARIES.MIN_EPISODES_SINCE));
 
 function buildSummaryPrompt(episodes, existingSummary = null) {
-    const context = episodes
+    const MAX_CHARS = 3000; // truncate input to keep Phi3 focused
+    let context = episodes
         .map(ep => `User: ${ep.user_message}\nAssistant: ${ep.ai_response}`)
         .join('\n\n');
 
+    // Truncate from the start if too long — keep the most recent exchanges
+    if (context.length > MAX_CHARS) {
+        context = context.slice(-MAX_CHARS);
+    }
+
     const instruction = existingSummary
-    ? `Update this summary to include the new conversation exchanges below. Output only the updated summary — do not continue the conversation.
+        ? `Update the summary below to include the new exchanges. Write 3-5 sentences in third person. Output only the updated summary text, nothing else.
 
 Previous summary:
 ${existingSummary}
 
-New exchanges to incorporate:`
+New exchanges:
-    : `Summarise this conversation. Output only the summary — do not continue the conversation or give recommendations.
+${context}`
+        : `Summarize the conversation below in 3-5 sentences. Write in third person. Output only the summary text, nothing else.
 
-Conversation:`;
+Conversation:
+${context}`;
 
     return [
-        '<|im_start|>system',
+        '<|user|>',
-        'You are a conversation summarisation assistant. You write concise, factual summaries for long-term memory storage. Output only the summary text — no preamble, no labels.<|im_end|>',
-        '<|im_start|>user',
         instruction,
-        '',
+        '<|end|>',
-        context,
+        '<|assistant|>',
-        '<|im_end|>',
-        '<|im_start|>assistant',
     ].join('\n');
 }
 
@@ -47,8 +51,8 @@ async function generateSummary(episodes, existingSummary = null) {
             prompt,
             stream: false,
             options: {
-                temperature: 0.3,   // slightly higher than entities — summaries benefit from some fluency
+                temperature: 0.2,   // slightly higher than entities — summaries benefit from some fluency
-                num_predict: 300,   // generous but bounded — keeps summaries from running long
+                num_predict: 200,   // generous but bounded — keeps summaries from running long
             },
         }),
     });