Notes/src/services/llm/formatters/ollama_formatter.ts
perf3ct 26b1b08129
tool calling is close to working
getting closer to calling tools...

we definitely need this

closer to tool execution...

agentic tool calling is...kind of working?
2025-04-08 19:15:01 +00:00

225 lines
9.7 KiB
TypeScript

import type { Message } from '../ai_interface.js';
import { BaseMessageFormatter } from './base_formatter.js';
import sanitizeHtml from 'sanitize-html';
import { PROVIDER_PROMPTS, FORMATTING_PROMPTS } from '../constants/llm_prompt_constants.js';
import { LLM_CONSTANTS } from '../constants/provider_constants.js';
import {
HTML_ALLOWED_TAGS,
HTML_ALLOWED_ATTRIBUTES,
OLLAMA_CLEANING,
FORMATTER_LOGS
} from '../constants/formatter_constants.js';
import log from '../../log.js';
/**
* Ollama-specific message formatter
* Handles the unique requirements of the Ollama API
*/
export class OllamaMessageFormatter extends BaseMessageFormatter {
/**
* Maximum recommended context length for Ollama
* Smaller than other providers due to Ollama's handling of context
*/
private static MAX_CONTEXT_LENGTH = LLM_CONSTANTS.CONTEXT_WINDOW.OLLAMA;
/**
* Format messages for the Ollama API
* @param messages Messages to format
* @param systemPrompt Optional system prompt to use
* @param context Optional context to include
* @param preserveSystemPrompt When true, preserves existing system messages rather than replacing them
*/
formatMessages(messages: Message[], systemPrompt?: string, context?: string, preserveSystemPrompt?: boolean): Message[] {
const formattedMessages: Message[] = [];
// Log the input messages with all their properties
log.info(`Ollama formatter received ${messages.length} messages`);
messages.forEach((msg, index) => {
const msgKeys = Object.keys(msg);
log.info(`Message ${index} - role: ${msg.role}, keys: ${msgKeys.join(', ')}, content length: ${msg.content.length}`);
// Log special properties if present
if (msg.tool_calls) {
log.info(`Message ${index} has ${msg.tool_calls.length} tool_calls`);
}
if (msg.tool_call_id) {
log.info(`Message ${index} has tool_call_id: ${msg.tool_call_id}`);
}
if (msg.name) {
log.info(`Message ${index} has name: ${msg.name}`);
}
});
// First identify user, system, and tool messages
const systemMessages = messages.filter(msg => msg.role === 'system');
const nonSystemMessages = messages.filter(msg => msg.role !== 'system');
// Determine if we should preserve the existing system message
if (preserveSystemPrompt && systemMessages.length > 0) {
// Preserve the existing system message
formattedMessages.push(systemMessages[0]);
log.info(`Preserving existing system message: ${systemMessages[0].content.substring(0, 50)}...`);
} else {
// Use provided systemPrompt or default
const basePrompt = systemPrompt || PROVIDER_PROMPTS.COMMON.DEFAULT_ASSISTANT_INTRO;
formattedMessages.push({
role: 'system',
content: basePrompt
});
log.info(`Using new system message: ${basePrompt.substring(0, 50)}...`);
}
// If we have context, inject it into the first user message
if (context && nonSystemMessages.length > 0) {
let injectedContext = false;
for (let i = 0; i < nonSystemMessages.length; i++) {
const msg = nonSystemMessages[i];
if (msg.role === 'user' && !injectedContext) {
// Simple context injection directly in the user's message
const cleanedContext = this.cleanContextContent(context);
log.info(`Injecting context (${cleanedContext.length} chars) into user message`);
const formattedContext = PROVIDER_PROMPTS.OLLAMA.CONTEXT_INJECTION(
cleanedContext,
msg.content
);
// Log what properties we're preserving
const msgKeys = Object.keys(msg);
const preservedKeys = msgKeys.filter(key => key !== 'role' && key !== 'content');
log.info(`Preserving additional properties in user message: ${preservedKeys.join(', ')}`);
// Create a new message with all original properties, but updated content
const newMessage = {
...msg, // Copy all properties
content: formattedContext // Override content with injected context
};
formattedMessages.push(newMessage);
log.info(`Created user message with context, final keys: ${Object.keys(newMessage).join(', ')}`);
injectedContext = true;
} else {
// For other messages, preserve all properties including any tool-related ones
log.info(`Preserving message with role ${msg.role}, keys: ${Object.keys(msg).join(', ')}`);
formattedMessages.push({
...msg // Copy all properties
});
}
}
} else {
// No context, just add all messages as-is
// Make sure to preserve all properties including tool_calls, tool_call_id, etc.
for (const msg of nonSystemMessages) {
log.info(`Adding message with role ${msg.role} without context injection, keys: ${Object.keys(msg).join(', ')}`);
formattedMessages.push({
...msg // Copy all properties
});
}
}
// Log the final formatted messages
log.info(`Ollama formatter produced ${formattedMessages.length} formatted messages`);
formattedMessages.forEach((msg, index) => {
const msgKeys = Object.keys(msg);
log.info(`Formatted message ${index} - role: ${msg.role}, keys: ${msgKeys.join(', ')}, content length: ${msg.content.length}`);
// Log special properties if present
if (msg.tool_calls) {
log.info(`Formatted message ${index} has ${msg.tool_calls.length} tool_calls`);
}
if (msg.tool_call_id) {
log.info(`Formatted message ${index} has tool_call_id: ${msg.tool_call_id}`);
}
if (msg.name) {
log.info(`Formatted message ${index} has name: ${msg.name}`);
}
});
return formattedMessages;
}
/**
* Clean up HTML and other problematic content before sending to Ollama
* Ollama needs a more aggressive cleaning than other models,
* but we want to preserve our XML tags for context
*/
override cleanContextContent(content: string): string {
if (!content) return '';
try {
// Store our XML tags so we can restore them after cleaning
const noteTagsRegex = /<\/?note>/g;
const notesTagsRegex = /<\/?notes>/g;
const queryTagsRegex = /<\/?query>[^<]*<\/query>/g;
// Capture tags to restore later
const noteTags = content.match(noteTagsRegex) || [];
const noteTagPositions: number[] = [];
let match;
const regex = /<\/?note>/g;
while ((match = regex.exec(content)) !== null) {
noteTagPositions.push(match.index);
}
// Remember the notes tags
const notesTagsMatch = content.match(notesTagsRegex) || [];
const notesTagPositions: number[] = [];
while ((match = notesTagsRegex.exec(content)) !== null) {
notesTagPositions.push(match.index);
}
// Remember the query tags
const queryTagsMatch = content.match(queryTagsRegex) || [];
// Temporarily replace XML tags with markers that won't be affected by sanitization
let modified = content
.replace(/<note>/g, '[NOTE_START]')
.replace(/<\/note>/g, '[NOTE_END]')
.replace(/<notes>/g, '[NOTES_START]')
.replace(/<\/notes>/g, '[NOTES_END]')
.replace(/<query>(.*?)<\/query>/g, '[QUERY]$1[/QUERY]');
// First use the parent class to do standard cleaning
let sanitized = super.cleanContextContent(modified);
// Then apply Ollama-specific aggressive cleaning
// Remove any remaining HTML using sanitizeHtml while keeping our markers
let plaintext = sanitizeHtml(sanitized, {
allowedTags: HTML_ALLOWED_TAGS.NONE,
allowedAttributes: HTML_ALLOWED_ATTRIBUTES.NONE,
textFilter: (text) => text
});
// Apply all Ollama-specific cleaning patterns
const ollamaPatterns = OLLAMA_CLEANING;
for (const pattern of Object.values(ollamaPatterns)) {
plaintext = plaintext.replace(pattern.pattern, pattern.replacement);
}
// Restore our XML tags
plaintext = plaintext
.replace(/\[NOTE_START\]/g, '<note>')
.replace(/\[NOTE_END\]/g, '</note>')
.replace(/\[NOTES_START\]/g, '<notes>')
.replace(/\[NOTES_END\]/g, '</notes>')
.replace(/\[QUERY\](.*?)\[\/QUERY\]/g, '<query>$1</query>');
return plaintext.trim();
} catch (error) {
console.error(FORMATTER_LOGS.ERROR.CONTEXT_CLEANING("Ollama"), error);
return content; // Return original if cleaning fails
}
}
/**
* Get the maximum recommended context length for Ollama
*/
getMaxContextLength(): number {
return OllamaMessageFormatter.MAX_CONTEXT_LENGTH;
}
}