clean up some shit

.env.example

@@ -11,3 +11,4 @@ OLLAMA_MODEL="" # Ollama model for responses e.g dolphin-mistral:latest
 FETCH_INTERVAL="" # interval for fetching new notifications from the instance, in milliseconds, recommend at least 15000
 RANDOM_POST_INTERVAL="" # interval for ad-hoc posts in milliseconds
 INSTANCE_BEARER_TOKEN="" # instance auth/bearer token (check the "verify_credentials" endpoint request headers in Chrome DevTools if on Soapbox)
+AD_HOC_POST_INSTRUCTIONS=""

src/main.ts

@@ -4,12 +4,10 @@ import {
   OllamaConfigOptions,
   OllamaChatRequest,
   OllamaChatResponse,
-  PostAncestorsForModel,
+  // PostAncestorsForModel,
 } from "../types.js";
-// import striptags from "striptags";
 import { PrismaClient } from "../generated/prisma/client.js";
 import {
-  // getInstanceEmojis,
   deleteNotification,
   getNotifications,
   getStatusContext,
@@ -19,8 +17,6 @@ import {
   isFromWhitelistedDomain,
   alreadyRespondedTo,
   recordPendingResponse,
-  // trimInputData,
-  // selectRandomEmoji,
   shouldContinue,
 } from "./util.js";
 
@@ -34,7 +30,7 @@ export const envConfig = {
     ? process.env.WHITELISTED_DOMAINS.split(",")
     : [process.env.PLEROMA_INSTANCE_DOMAIN],
   ollamaUrl: process.env.OLLAMA_URL || "",
-  ollamaSystemPrompt: process.env.OLLAMA_SYSTEM_PROMPT,
+  ollamaSystemPrompt: process.env.OLLAMA_SYSTEM_PROMPT || "",
   ollamaModel: process.env.OLLAMA_MODEL || "",
   fetchInterval: process.env.FETCH_INTERVAL
     ? parseInt(process.env.FETCH_INTERVAL)
@@ -45,14 +41,18 @@ export const envConfig = {
     : 3600000,
   botAccountId: process.env.PLEROMA_ACCOUNT_ID,
   replyWithContext: process.env.REPLY_WITH_CONTEXT === "true" ? true : false,
+  adHocPostInstructions: process.env.AD_HOC_POST_INSTRUCTIONS
+    ? process.env.AD_HOC_POST_INSTRUCTIONS
+    : "Say something.",
 };
 
 const ollamaConfig: OllamaConfigOptions = {
-  temperature: 0.9,
-  top_p: 0.85,
-  top_k: 60,
-  num_ctx: 16384, // maximum context window for Llama 3.1
-  repeat_penalty: 1.1,
+  temperature: 0.85, // Increased from 0.6 - more creative and varied
+  top_p: 0.9, // Slightly increased for more diverse responses
+  top_k: 40,
+  num_ctx: 16384,
+  repeat_penalty: 1.1, // Reduced from 1.15 - less mechanical
+  // stop: ['<|im_end|>', '\n\n']
 };
 
 // this could be helpful
@@ -68,75 +68,128 @@ const generateOllamaRequest = async (
     ollamaUrl,
     replyWithContext,
   } = envConfig;
+
+  let shouldDeleteNotification = false;
+
   try {
-    if (shouldContinue(notification)) {
-      if (whitelistOnly && !isFromWhitelistedDomain(notification)) {
-        await deleteNotification(notification);
-        return;
-      }
-      if (await alreadyRespondedTo(notification)) {
-        return;
-      }
-      await recordPendingResponse(notification);
-      await storeUserData(notification);
-      let conversationHistory: PostAncestorsForModel[] = [];
-      if (replyWithContext) {
-        const contextPosts = await getStatusContext(notification.status.id);
-        if (!contextPosts?.ancestors || !contextPosts) {
-          throw new Error(`Unable to obtain post context ancestors.`);
-        }
-        conversationHistory = contextPosts.ancestors.map((ancestor) => {
-          const mentions = ancestor.mentions.map((mention) => mention.acct);
-          return {
-            account_fqn: ancestor.account.fqn,
-            mentions,
-            plaintext_content: ancestor.pleroma.content["text/plain"],
-          };
-        });
-        // console.log(conversationHistory);
-      }
-
-      // Simplified user message (remove [/INST] as it's not needed for Llama 3)
-      const userMessage = `${notification.status.account.fqn} says to you: \"${notification.status.pleroma.content["text/plain"]}\".`;
-
-      let systemContent = ollamaSystemPrompt;
-      if (replyWithContext) {
-        // Simplified context instructions (avoid heavy JSON; summarize for clarity)
-        systemContent = `${ollamaSystemPrompt}\n\nPrevious conversation context:\n${conversationHistory
-          .map(
-            (post) =>
-              `${post.account_fqn} (said to ${post.mentions.join(", ")}): ${
-                post.plaintext_content
-              }`
-          )
-          .join(
-            "\n"
-          )}\nReply to the user who addressed you (you are Lexi, also known as nice-ai or nice-ai@nicecrew.digital). Examine the context of the entire conversation and make references to topics or information where appropriate. Prefix usernames with '@' when addressing them. Assume if there is no domain in the username, the domain is @nicecrew.digital (for example @matty would be @matty@nicecrew.digital)`;
-      }
-
-      // Switch to chat request format (messages array auto-handles Llama 3 template)
-      const ollamaRequestBody: OllamaChatRequest = {
-        model: ollamaModel,
-        messages: [
-          { role: "system", content: systemContent as string },
-          { role: "user", content: userMessage },
-        ],
-        stream: false,
-        options: ollamaConfig,
-      };
-
-      // Change endpoint to /api/chat
-      const response = await fetch(`${ollamaUrl}/api/chat`, {
-        method: "POST",
-        body: JSON.stringify(ollamaRequestBody),
-      });
-      const ollamaResponse: OllamaChatResponse = await response.json();
-
-      await storePromptData(notification, ollamaResponse);
-      return ollamaResponse;
+    if (!shouldContinue(notification)) {
+      shouldDeleteNotification = true;
+      return;
     }
+
+    if (whitelistOnly && !isFromWhitelistedDomain(notification)) {
+      shouldDeleteNotification = true;
+      return;
+    }
+
+    if (await alreadyRespondedTo(notification)) {
+      shouldDeleteNotification = true;
+      return;
+    }
+
+    await recordPendingResponse(notification);
+    await storeUserData(notification);
+
+    let conversationContext = "";
+    if (replyWithContext) {
+      const contextPosts = await getStatusContext(notification.status.id);
+      if (!contextPosts?.ancestors) {
+        throw new Error(`Unable to obtain post context ancestors.`);
+      }
+
+      // Build a human-readable conversation thread
+      const allPosts = [...contextPosts.ancestors];
+
+      // Include descendants (follow-up posts) if available
+      if (contextPosts.descendents && contextPosts.descendents.length > 0) {
+        allPosts.push(...contextPosts.descendents);
+      }
+
+      if (allPosts.length > 0) {
+        const conversationLines = allPosts.map((post) => {
+          const author = post.account.fqn;
+          const content = post.pleroma.content["text/plain"];
+          const replyingTo = post.in_reply_to_account_id
+            ? ` (replying to another message)`
+            : "";
+          return `[@${author}${replyingTo}]: ${content}`;
+        });
+
+        conversationContext = `
+Previous conversation thread:
+${conversationLines.join("\n\n")}
+---
+`;
+      }
+    }
+
+    const userMessage = notification.status.pleroma.content["text/plain"];
+    const originalAuthor = notification.account.fqn;
+
+    let systemContent = ollamaSystemPrompt;
+    if (replyWithContext && conversationContext) {
+      systemContent = `${ollamaSystemPrompt}
+
+${conversationContext}
+Current message from @${originalAuthor}:
+"${userMessage}"
+
+Instructions:
+- You are replying to @${originalAuthor}
+- Address them directly if appropriate
+- Use markdown formatting and emojis sparingly`;
+    }
+
+    const ollamaRequestBody: OllamaChatRequest = {
+      model: ollamaModel,
+      messages: [
+        { role: "system", content: systemContent },
+        { role: "user", content: userMessage },
+      ],
+      stream: false,
+      options: {
+        ...ollamaConfig,
+        stop: ["</s>", "[INST]"], // Mistral 0.3 stop tokens
+      },
+    };
+
+    console.log(
+      `Generating response for notification ${notification.id} from @${originalAuthor}`
+    );
+
+    // Change endpoint to /api/chat
+    const response = await fetch(`${ollamaUrl}/api/chat`, {
+      method: "POST",
+      body: JSON.stringify(ollamaRequestBody),
+    });
+
+    if (!response.ok) {
+      throw new Error(`Ollama API request failed: ${response.statusText}`);
+    }
+
+    const ollamaResponse: OllamaChatResponse = await response.json();
+
+    await storePromptData(notification, ollamaResponse);
+    return ollamaResponse;
   } catch (error: any) {
-    throw new Error(error.message);
+    console.error(
+      `Error in generateOllamaRequest for notification ${notification.id}:`,
+      error.message
+    );
+    // Delete notification on error to prevent retry loops
+    shouldDeleteNotification = true;
+    throw error;
+  } finally {
+    if (shouldDeleteNotification) {
+      try {
+        await deleteNotification(notification);
+      } catch (deleteError: any) {
+        console.error(
+          `Failed to delete notification ${notification.id}:`,
+          deleteError.message
+        );
+      }
+    }
   }
 };
 
@@ -145,27 +198,28 @@ const postReplyToStatus = async (
   ollamaResponseBody: OllamaChatResponse
 ) => {
   const { pleromaInstanceUrl, bearerToken } = envConfig;
-  // const emojiList = await getInstanceEmojis();
-  // let randomEmoji;
-  // if (emojiList) {
-  //   randomEmoji = selectRandomEmoji(emojiList);
-  // }
+
   try {
-    let mentions: string[];
+    // Only mention the original author who triggered the bot
+    const originalAuthor = notification.account.acct;
+    console.log(
+      `Replying to: @${originalAuthor} (status ID: ${notification.status.id})`
+    );
+
+    // Sanitize LLM output - remove any stray Mistral special tokens
+    let sanitizedContent = ollamaResponseBody.message.content
+      .replace(/<\/s>/g, "") // Remove EOS token if it appears
+      .replace(/\[INST\]/g, "") // Remove instruction start token
+      .replace(/\[\/INST\]/g, "") // Remove instruction end token
+      .replace(/<s>/g, "") // Remove BOS token if it appears
+      .trim();
+
     const statusBody: NewStatusBody = {
       content_type: "text/markdown",
-      status: `${ollamaResponseBody.message.content}`,
+      status: sanitizedContent,
       in_reply_to_id: notification.status.id,
+      to: [originalAuthor], // Only send to the person who mentioned the bot
     };
-    if (
-      notification.status.mentions &&
-      notification.status.mentions.length > 0
-    ) {
-      mentions = notification.status.mentions.map((mention) => {
-        return mention.acct;
-      });
-      statusBody.to = mentions;
-    }
 
     const response = await fetch(`${pleromaInstanceUrl}/api/v1/statuses`, {
       method: "POST",
@@ -180,9 +234,23 @@ const postReplyToStatus = async (
       throw new Error(`New status request failed: ${response.statusText}`);
     }
 
-    await deleteNotification(notification);
+    console.log(`Successfully posted reply to @${originalAuthor}`);
   } catch (error: any) {
-    throw new Error(error.message);
+    console.error(
+      `Error posting reply for notification ${notification.id}:`,
+      error.message
+    );
+    throw error;
+  } finally {
+    // Always try to delete the notification, even if posting failed
+    try {
+      await deleteNotification(notification);
+    } catch (deleteError: any) {
+      console.error(
+        `Failed to delete notification ${notification.id}:`,
+        deleteError.message
+      );
+    }
   }
 };
 
@@ -193,15 +261,22 @@ const createTimelinePost = async () => {
     ollamaSystemPrompt,
     ollamaUrl,
     pleromaInstanceUrl,
+    adHocPostInstructions,
   } = envConfig;
   const ollamaRequestBody: OllamaChatRequest = {
     model: ollamaModel,
     messages: [
-      { role: "system", content: ollamaSystemPrompt as string },
-      { role: "user", content: "Say something random." },
+      { role: "system", content: ollamaSystemPrompt },
+      {
+        role: "user",
+        content: adHocPostInstructions,
+      },
     ],
     stream: false,
-    options: ollamaConfig,
+    options: {
+      ...ollamaConfig,
+      stop: ["</s>", "[INST]"], // Mistral 0.3 stop tokens
+    },
   };
   try {
     const response = await fetch(`${ollamaUrl}/api/chat`, {
@@ -244,18 +319,21 @@ const beginFetchCycle = async () => {
   setInterval(async () => {
     notifications = await getNotifications();
     if (notifications.length > 0) {
-      await Promise.all(
-        notifications.map(async (notification) => {
-          try {
-            const ollamaResponse = await generateOllamaRequest(notification);
-            if (ollamaResponse) {
-              postReplyToStatus(notification, ollamaResponse);
-            }
-          } catch (error: any) {
-            throw new Error(error.message);
+      // Process notifications sequentially to avoid race conditions
+      for (const notification of notifications) {
+        try {
+          const ollamaResponse = await generateOllamaRequest(notification);
+          if (ollamaResponse) {
+            await postReplyToStatus(notification, ollamaResponse);
           }
-        })
-      );
+        } catch (error: any) {
+          console.error(
+            `Error processing notification ${notification.id}:`,
+            error.message
+          );
+          // Continue processing other notifications even if one fails
+        }
+      }
     }
   }, envConfig.fetchInterval); // lower intervals may cause the bot to respond multiple times to the same message, but we try to mitigate this with the deleteNotification function
 };
@@ -277,6 +355,11 @@ console.log(
     envConfig.fetchInterval / 1000
   } seconds.`
 );
+console.log(
+  `Making ad-hoc post to ${envConfig.pleromaInstanceDomain}, every ${
+    envConfig.adHocPostInterval / 1000 / 60
+  } minutes.`
+);
 console.log(
   `Accepting prompts from: ${envConfig.whitelistedDomains.join(", ")}`
 );
@@ -288,7 +371,4 @@ console.log(
 console.log(`System prompt: ${envConfig.ollamaSystemPrompt}`);
 
 await beginFetchCycle();
-// setInterval(async () => {
-//   createTimelinePost();
-// }, 10000);
 await beginStatusPostInterval();