Better memory de-duplication

package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@ztimson/ai-utils",
-	"version": "0.8.1",
+	"version": "0.8.5",
 	"description": "AI Utility library",
 	"author": "Zak Timson",
 	"license": "MIT",

src/llm.ts

@@ -44,8 +44,6 @@ export type LLMMemory = {
 	fact: string;
 	/** Owner and fact embedding vector */
 	embeddings: [number[], number[]];
-	/** Creation time */
-	timestamp: Date;
 }
 
 export type LLMRequest = {
@@ -95,71 +93,85 @@ class LLM {
 	 * @returns {{abort: () => void, response: Promise<string>}} Function to abort response and chat history
 	 */
 	ask(message: string, options: LLMRequest = {}): AbortablePromise<string> {
+		options = <any>{
+			system: '',
+			temperature: 0.8,
+			...this.ai.options.llm,
+			models: undefined,
+			history: [],
+			...options,
+		}
 		const m = options.model || this.defaultModel;
 		if(!this.models[m]) throw new Error(`Model does not exist: ${m}`);
 		let abort = () => {};
 		return Object.assign(new Promise<string>(async res => {
 			if(!options.history) options.history = [];
-			// If memories were passed, find any relivant ones and add a tool for ADHOC lookups
+			// If memories were passed, find any relevant ones and add a tool for ADHOC lookups
 			if(options.memory) {
-				options.system = (options.system || '') + '\nYou have passive persistent memory never make any mention of your memory capabilities and what you can/cannot remember\n';
+				const search = async (query?: string | null, subject?: string | null, limit = 10) => {
-				const search = async (query?: string | null, subject?: string | null, limit = 50) => {
 					const [o, q] = await Promise.all([
 						subject ? this.embedding(subject) : Promise.resolve(null),
 						query ? this.embedding(query) : Promise.resolve(null),
 					]);
-					return (options.memory || [])
+					return (options.memory || []).map(m => {
-						.map(m => ({...m, score: o ? this.cosineSimilarity(m.embeddings[0], o[0].embedding) : 1}))
+						const score = (o ? this.cosineSimilarity(m.embeddings[0], o[0].embedding) : 0)
-						.filter((m: any) => m.score >= 0.8)
+							+ (q ? this.cosineSimilarity(m.embeddings[1], q[0].embedding) : 0);
-						.map((m: any) => ({...m, score: q ? this.cosineSimilarity(m.embeddings[1], q[0].embedding) : m.score}))
+						return {...m, score};
-						.filter((m: any) => m.score >= 0.2)
+					}).toSorted((a: any, b: any) => a.score - b.score).slice(0, limit);
-						.toSorted((a: any, b: any) => a.score - b.score)
-						.slice(0, limit);
 				}
 
+				options.system += '\nYou have RAG memory and will be given the top_k closest memories regarding the users query. Save anything new you have learned worth remembering from the user message using the remember tool and feel free to recall memories manually.\n';
 				const relevant = await search(message);
-				if(relevant.length) options.history.push({role: 'assistant', content: 'Things I remembered:\n' + relevant.map(m => `${m.owner}: ${m.fact}`).join('\n')});
+				if(relevant.length) options.history.push({role: 'tool', name: 'recall', id: 'auto_recall_' + Math.random().toString(), args: {}, content: 'Things I remembered:\n' + relevant.map(m => `${m.owner}: ${m.fact}`).join('\n')});
-				options.tools = [...options.tools || [], {
+				options.tools = [{
-					name: 'read_memory',
+					name: 'recall',
-					description: 'Check your long-term memory for more information',
+					description: 'Recall the closest memories you have regarding a query using RAG',
 					args: {
 						subject: {type: 'string', description: 'Find information by a subject topic, can be used with or without query argument'},
 						query: {type: 'string', description: 'Search memory based on a query, can be used with or without subject argument'},
-						limit: {type: 'number', description: 'Result limit, default 5'},
+						topK: {type: 'number', description: 'Result limit, default 5'},
 					},
 					fn: (args) => {
 						if(!args.subject && !args.query) throw new Error('Either a subject or query argument is required');
-						return search(args.query, args.subject, args.limit || 5);
+						return search(args.query, args.subject, args.topK);
 					}
-				}];
+				}, {
+					name: 'remember',
+					description: 'Store important facts user shares for future recall',
+					args: {
+						owner: {type: 'string', description: 'Subject/person this fact is about'},
+						fact: {type: 'string', description: 'The information to remember'}
+					},
+					fn: async (args) => {
+						if(!options.memory) return;
+						const e = await Promise.all([
+							this.embedding(args.owner),
+							this.embedding(`${args.owner}: ${args.fact}`)
+						]);
+						const newMem = {owner: args.owner, fact: args.fact, embeddings: <any>[e[0][0].embedding, e[1][0].embedding]};
+						options.memory.splice(0, options.memory.length, ...[
+							...options.memory.filter(m => {
+								return !(this.cosineSimilarity(newMem.embeddings[0], m.embeddings[0]) >= 0.9 && this.cosineSimilarity(newMem.embeddings[1], m.embeddings[1]) >= 0.8);
+							}),
+							newMem
+						]);
+						return 'Remembered!';
+					}
+				}, ...options.tools || []];
 			}
 
 			// Ask
 			const resp = await this.models[m].ask(message, options);
 
-			// Remove any memory calls
+			// Remove any memory calls from history
-			if(options.memory) {
+			if(options.memory) options.history.splice(0, options.history.length, ...options.history.filter(h => h.role != 'tool' || (h.name != 'recall' && h.name != 'remember')));
-				const i = options.history?.findIndex((h: any) => h.role == 'assistant' && h.content.startsWith('Things I remembered:'));
+
-				if(i != null && i >= 0) options.history?.splice(i, 1);
+			// Compress message history
+			if(options.compress) {
+				const compressed = await this.ai.language.compressHistory(options.history, options.compress.max, options.compress.min, options);
+				options.history.splice(0, options.history.length, ...compressed);
 			}
 
-			// Handle compression and memory extraction
-			if(options.compress || options.memory) {
-				let compressed: any = null;
-				if(options.compress) {
-					compressed = await this.ai.language.compressHistory(options.history, options.compress.max, options.compress.min, options);
-					options.history.splice(0, options.history.length, ...compressed.history);
-				} else {
-					const i = options.history?.findLastIndex(m => m.role == 'user') ?? -1;
-					compressed = await this.ai.language.compressHistory(i != -1 ? options.history.slice(i) : options.history, 0, 0, options);
-				}
-				if(options.memory) {
-					const updated = options.memory
-						.filter(m => !compressed.memory.some(m2 => this.cosineSimilarity(m.embeddings[1], m2.embeddings[1]) > 0.8))
-						.concat(compressed.memory);
-					options.memory.splice(0, options.memory.length, ...updated);
-				}
-			}
 			return res(resp);
 		}), {abort});
 	}
@@ -181,32 +193,24 @@ class LLM {
 	 * @param {LLMRequest} options LLM options
 	 * @returns {Promise<LLMMessage[]>} New chat history will summary at index 0
 	 */
-	async compressHistory(history: LLMMessage[], max: number, min: number, options?: LLMRequest): Promise<{history: LLMMessage[], memory: LLMMemory[]}> {
+	async compressHistory(history: LLMMessage[], max: number, min: number, options?: LLMRequest): Promise<LLMMessage[]> {
-		if(this.estimateTokens(history) < max) return {history, memory: []};
+		if(this.estimateTokens(history) < max) return history;
 		let keep = 0, tokens = 0;
 		for(let m of history.toReversed()) {
 			tokens += this.estimateTokens(m.content);
 			if(tokens < min) keep++;
 			else break;
 		}
-		if(history.length <= keep) return {history, memory: []};
+		if(history.length <= keep) return history;
 		const system = history[0].role == 'system' ? history[0] : null,
 			recent = keep == 0 ? [] : history.slice(-keep),
 			process = (keep == 0 ? history : history.slice(0, -keep)).filter(h => h.role === 'assistant' || h.role === 'user');
 
-		const summary: any = await this.json(process.map(m => `${m.role}: ${m.content}`).join('\n\n'), '{summary: string, facts: [[subject, fact]]}', {
+		const summary: any = await this.summarize(process.map(m => `[${m.role}]: ${m.content}`).join('\n\n'), 500, options);
-			system: 'Create the smallest summary possible, no more than 500 tokens. Create a list of NEW facts (split by subject [pro]noun and fact) about what you learned from this conversation that you didn\'t already know or get from a tool call or system prompt. Focus only on new information about people, topics, or facts. Avoid generating facts about the AI.',
+		const d = Date.now();
-			model: options?.model,
+		const h = [{role: <any>'tool', name: 'summary', id: `summary_` + d, args: {}, content: `Conversation Summary: ${summary?.summary}`, timestamp: d}, ...recent];
-			temperature: options?.temperature || 0.3
-		});
-		const timestamp = new Date();
-		const memory = await Promise.all((summary?.facts || [])?.map(async ([owner, fact]: [string, string]) => {
-			const e = await Promise.all([this.embedding(owner), this.embedding(`${owner}: ${fact}`)]);
-			return {owner, fact, embeddings: [e[0][0].embedding, e[1][0].embedding], timestamp};
-		}));
-		const h = [{role: 'assistant', content: `Conversation Summary: ${summary?.summary}`, timestamp: Date.now()}, ...recent];
 		if(system) h.splice(0, 0, system);
-		return {history: <any>h, memory};
+		return h;
 	}
 
 	/**
@@ -243,7 +247,7 @@ class LLM {
 				return `${p}: ${Array.isArray(value) ? value.join(', ') : value}`;
 			});
 		};
-		const lines = typeof target === 'object' ? objString(target) : target.split('\n');
+		const lines = typeof target === 'object' ? objString(target) : target.toString().split('\n');
 		const tokens = lines.flatMap(l => [...l.split(/\s+/).filter(Boolean), '\n']);
 		const chunks: string[] = [];
 		for(let i = 0; i < tokens.length;) {
@@ -366,8 +370,8 @@ class LLM {
 	 * @param options LLM request options
 	 * @returns {Promise<string>} Summary
 	 */
-	summarize(text: string, tokens: number, options?: LLMRequest): Promise<string | null> {
+	summarize(text: string, tokens: number = 500, options?: LLMRequest): Promise<string | null> {
-		return this.ask(text, {system: `Generate a brief summary <= ${tokens} tokens. Output nothing else`, temperature: 0.3, ...options});
+		return this.ask(text, {system: `Generate the shortest summary possible <= ${tokens} tokens. Output nothing else`, temperature: 0.3, ...options});
 	}
 }
 

src/open-ai.ts

@@ -11,7 +11,7 @@ export class OpenAi extends LLMProvider {
 		super();
 		this.client = new openAI(clean({
 			baseURL: host,
-			apiKey: token
+			apiKey: token || host ? 'ignored' : undefined
 		}));
 	}
 
@@ -67,7 +67,10 @@ export class OpenAi extends LLMProvider {
 	ask(message: string, options: LLMRequest = {}): AbortablePromise<string> {
 		const controller = new AbortController();
 		return Object.assign(new Promise<any>(async (res, rej) => {
-			if(options.system && options.history?.[0]?.role != 'system') options.history?.splice(0, 0, {role: 'system', content: options.system, timestamp: Date.now()});
+			if(options.system) {
+				if(options.history?.[0]?.role != 'system') options.history?.splice(0, 0, {role: 'system', content: options.system, timestamp: Date.now()});
+				else options.history[0].content = options.system;
+			}
 			let history = this.fromStandard([...options.history || [], {role: 'user', content: message, timestamp: Date.now()}]);
 			const tools = options.tools || this.ai.options.llm?.tools || [];
 			const requestParams: any = {
@@ -100,15 +103,37 @@ export class OpenAi extends LLMProvider {
 				if(options.stream) {
 					if(!isFirstMessage) options.stream({text: '\n\n'});
 					else isFirstMessage = false;
-					resp.choices = [{message: {content: '', tool_calls: []}}];
+					resp.choices = [{message: {role: 'assistant', content: '', tool_calls: []}}];
 					for await (const chunk of resp) {
 						if(controller.signal.aborted) break;
 						if(chunk.choices[0].delta.content) {
 							resp.choices[0].message.content += chunk.choices[0].delta.content;
 							options.stream({text: chunk.choices[0].delta.content});
 						}
+
 						if(chunk.choices[0].delta.tool_calls) {
-							resp.choices[0].message.tool_calls = chunk.choices[0].delta.tool_calls;
+							for(const deltaTC of chunk.choices[0].delta.tool_calls) {
+								const existing = resp.choices[0].message.tool_calls.find(tc => tc.index === deltaTC.index);
+								if(existing) {
+									if(deltaTC.id) existing.id = deltaTC.id;
+									if(deltaTC.type) existing.type = deltaTC.type;
+									if(deltaTC.function) {
+										if(!existing.function) existing.function = {};
+										if(deltaTC.function.name) existing.function.name = deltaTC.function.name;
+										if(deltaTC.function.arguments) existing.function.arguments = (existing.function.arguments || '') + deltaTC.function.arguments;
+									}
+								} else {
+									resp.choices[0].message.tool_calls.push({
+										index: deltaTC.index,
+										id: deltaTC.id || '',
+										type: deltaTC.type || 'function',
+										function: {
+											name: deltaTC.function?.name || '',
+											arguments: deltaTC.function?.arguments || ''
+										}
+									});
+								}
+							}
 						}
 					}
 				}