@inferrlm/react-native-mlx 0.4.2-alpha.0 → 0.4.2-alpha.10

package/ios/Sources/HybridLLM.swift

@@ -29,7 +29,11 @@ class HybridLLM: HybridLLMSpec {
     var modelId: String = ""
     var debug: Bool = false
     var systemPrompt: String = "You are a helpful assistant."
+    var maxTokens: Int = 2048
+    var temperature: Float = 0.7
+    var enableThinking: Bool = true
     var additionalContext: LLMMessage = LLMMessage()
+    private var lastInputContainedThinkTag = false
 
     private func log(_ message: String) {
         if debug {
@@ -72,16 +76,27 @@ class HybridLLM: HybridLLMSpec {
               let json = try? JSONSerialization.jsonObject(with: data) as? [String: Any]
         else { return [] }
 
+        var allIds = Set<Int>()
+
         if let ids = extractEosIds(from: json) {
-            return ids
+            allIds.formUnion(ids)
+        }
+
+        for key in ["text_config", "language_config", "llm_config"] {
+            if let nested = json[key] as? [String: Any],
+               let ids = extractEosIds(from: nested) {
+                allIds.formUnion(ids)
+            }
         }
 
-        if let textConfig = json["text_config"] as? [String: Any],
-           let ids = extractEosIds(from: textConfig) {
-            return ids
+        let genConfigURL = modelDir.appendingPathComponent("generation_config.json")
+        if let genData = try? Data(contentsOf: genConfigURL),
+           let genJson = try? JSONSerialization.jsonObject(with: genData) as? [String: Any],
+           let ids = extractEosIds(from: genJson) {
+            allIds = ids
         }
 
-        return []
+        return allIds
     }
 
     private func extractEosIds(from dict: [String: Any]) -> Set<Int>? {
@@ -157,16 +172,27 @@ class HybridLLM: HybridLLMSpec {
                  mlx-swift-lm only reads top-level eos_token_id from config.json.
                  Models like Qwen3.5 nest it inside text_config, leaving the stop
                  set empty. Parse it ourselves and patch the container.
+                 Also add common chat stop tokens as extraEOSTokens.
                 */
                 let containerEos = await loadedContainer.configuration.eosTokenIds
-                if containerEos.isEmpty {
-                    let parsed = self.parseEosTokenIds(from: modelDir)
-                    if !parsed.isEmpty {
-                        log("Patching eosTokenIds from config: \(parsed)")
-                        await loadedContainer.update { ctx in
+                let containerExtra = await loadedContainer.configuration.extraEOSTokens
+                log("EOS state after load - ids: \(containerEos), extra: \(containerExtra)")
+
+                let parsed = self.parseEosTokenIds(from: modelDir)
+                let chatStopTokens: Set<String> = ["<|endoftext|>", "<|im_end|>", "<|im_start|>"]
+                let needsIdPatch = containerEos.isEmpty && !parsed.isEmpty
+                let missingExtra = chatStopTokens.subtracting(containerExtra)
+
+                if needsIdPatch || !missingExtra.isEmpty {
+                    await loadedContainer.update { ctx in
+                        if needsIdPatch {
                             ctx.configuration.eosTokenIds = parsed
                         }
+                        ctx.configuration.extraEOSTokens.formUnion(chatStopTokens)
                     }
+                    let updated = await loadedContainer.configuration.eosTokenIds
+                    let updatedExtra = await loadedContainer.configuration.extraEOSTokens
+                    log("EOS patched - ids: \(updated), extra: \(updatedExtra)")
                 }
 
                 let memoryAfterContainer = self.getMemoryUsage()
@@ -208,10 +234,6 @@ class HybridLLM: HybridLLMSpec {
         }
 
         return Promise.async { [self] in
-            if self.manageHistory {
-                self.messageHistory.append(LLMMessage(role: "user", content: prompt))
-            }
-
             let task = Task<String, Error> {
                 log("Generating response for: \(prompt.prefix(50))...")
                 let result = try await session.respond(to: prompt)
@@ -225,6 +247,7 @@ class HybridLLM: HybridLLMSpec {
             let result = try await task.value
 
             if self.manageHistory {
+                self.messageHistory.append(LLMMessage(role: "user", content: prompt))
                 self.messageHistory.append(LLMMessage(role: "assistant", content: result))
             }
 
@@ -243,16 +266,21 @@ class HybridLLM: HybridLLMSpec {
             throw LLMError.notLoaded
         }
 
-        return Promise.async { [self] in
-            if self.manageHistory {
-                self.messageHistory.append(LLMMessage(role: "user", content: prompt))
-            }
+        if let prev = currentTask {
+            log("stream_cancelling_previous")
+            prev.cancel()
+            currentTask = nil
+        }
 
+        return Promise.async { [self] in
             let task = Task<String, Error> {
                 let startTime = Date()
                 var firstTokenTime: Date?
                 var tokenCount = 0
 
+                log("stream_start prompt=\(prompt.count)chars history=\(self.messageHistory.count) manageHistory=\(self.manageHistory) maxTokens=\(self.maxTokens) temperature=\(self.temperature)")
+                log("stream_prompt: \(prompt)")
+
                 let result = try await self.performGeneration(
                     container: container,
                     prompt: prompt,
@@ -281,7 +309,8 @@ class HybridLLM: HybridLLMSpec {
                     toolExecutionTime: 0
                 )
 
-                log("Stream complete - \(tokenCount) tokens, \(String(format: "%.1f", tokensPerSecond)) tokens/s")
+                log("stream_done tokens=\(tokenCount) tps=\(String(format: "%.1f", tokensPerSecond)) result=\(result.count)chars")
+                log("stream_result: \(result)")
                 return result
             }
 
@@ -291,7 +320,9 @@ class HybridLLM: HybridLLMSpec {
             let result = try await task.value
 
             if self.manageHistory {
+                self.messageHistory.append(LLMMessage(role: "user", content: prompt))
                 self.messageHistory.append(LLMMessage(role: "assistant", content: result))
+                log("stream_history_updated count=\(self.messageHistory.count)")
             }
 
             return result
@@ -306,11 +337,13 @@ class HybridLLM: HybridLLMSpec {
             throw LLMError.notLoaded
         }
 
-        return Promise.async { [self] in
-            if self.manageHistory {
-                self.messageHistory.append(LLMMessage(role: "user", content: prompt))
-            }
+        if let prev = currentTask {
+            log("streamWithEvents_cancelling_previous")
+            prev.cancel()
+            currentTask = nil
+        }
 
+        return Promise.async { [self] in
             let task = Task<String, Error> {
                 let startTime = Date()
                 var firstTokenTime: Date?
@@ -367,6 +400,7 @@ class HybridLLM: HybridLLMSpec {
             let result = try await task.value
 
             if self.manageHistory {
+                self.messageHistory.append(LLMMessage(role: "user", content: prompt))
                 self.messageHistory.append(LLMMessage(role: "assistant", content: result))
             }
 
@@ -374,6 +408,8 @@ class HybridLLM: HybridLLMSpec {
         }
     }
 
+    private static let fallbackTemplate = "{%- for message in messages %}{{'<|im_start|>' + message['role'] + '\\n' + message['content'] + '<|im_end|>' + '\\n'}}{%- endfor %}{%- if add_generation_prompt %}{%- if enable_thinking is defined and enable_thinking is true %}{{ '<|im_start|>assistant\\n<think>\\n' }}{%- else %}{{ '<|im_start|>assistant\\n' }}{%- endif %}{%- endif %}"
+
     private func buildChatMessages(
         prompt: String,
         toolResults: [String]?,
@@ -381,11 +417,14 @@ class HybridLLM: HybridLLMSpec {
     ) -> [Chat.Message] {
         var chat: [Chat.Message] = []
 
+        log("build_chat depth=\(depth) history=\(self.messageHistory.count) prompt=\(prompt.count)chars")
+
         if !self.systemPrompt.isEmpty {
             chat.append(.system(self.systemPrompt))
+            log("  [system] \(self.systemPrompt)")
         }
 
-        for msg in self.messageHistory {
+        for (i, msg) in self.messageHistory.enumerated() {
             switch msg.role {
             case "user": chat.append(.user(msg.content))
             case "assistant": chat.append(.assistant(msg.content))
@@ -393,21 +432,71 @@ class HybridLLM: HybridLLMSpec {
             case "tool": chat.append(.tool(msg.content))
             default: break
             }
+            log("  [\(i):\(msg.role)] \(msg.content)")
         }
 
         if depth == 0 {
             chat.append(.user(prompt))
+            log("  [prompt] \(prompt)")
         }
 
         if let toolResults {
-            for result in toolResults {
+            for (i, result) in toolResults.enumerated() {
                 chat.append(.tool(result))
+                log("  [tool_result_\(i)] \(result)")
             }
         }
 
+        log("chat_built total=\(chat.count) messages")
         return chat
     }
 
+    private func prepareInput(
+        container: ModelContainer,
+        chat: [Chat.Message]
+    ) async throws -> LMInput {
+        let tools = !self.toolSchemas.isEmpty ? self.toolSchemas : nil
+        let thinkingEnabled = self.enableThinking
+        let additionalCtx: [String: any Sendable] = ["enable_thinking": thinkingEnabled]
+
+        let messages: [[String: any Sendable]] = chat.map {
+            ["role": $0.role.rawValue, "content": $0.content]
+        }
+
+        let tokens: [Int] = try await container.perform { (context: ModelContext) in
+            do {
+                let result = try context.tokenizer.applyChatTemplate(
+                    messages: messages,
+                    tools: tools,
+                    additionalContext: additionalCtx
+                )
+                self.log("template_applied token_count=\(result.count)")
+                let decoded = context.tokenizer.decode(tokens: Array(result.suffix(60)))
+                self.log("input_tail_decoded: \(decoded)")
+                self.lastInputContainedThinkTag = decoded.contains("<think>")
+                return result
+            } catch {
+                self.log("template_error: \(error), retrying with fallback")
+                let result = try context.tokenizer.applyChatTemplate(
+                    messages: messages,
+                    chatTemplate: .literal(HybridLLM.fallbackTemplate),
+                    addGenerationPrompt: true,
+                    truncation: false,
+                    maxLength: nil,
+                    tools: nil,
+                    additionalContext: additionalCtx
+                )
+                self.log("fallback_template_applied token_count=\(result.count)")
+                let decoded = context.tokenizer.decode(tokens: Array(result.suffix(60)))
+                self.log("fallback_input_tail_decoded: \(decoded)")
+                self.lastInputContainedThinkTag = decoded.contains("<think>")
+                return result
+            }
+        }
+
+        return LMInput(tokens: MLXArray(tokens))
+    }
+
     private func executeToolCall(
         tool: ToolDefinition,
         argsDict: [String: Any]
@@ -438,17 +527,34 @@ class HybridLLM: HybridLLMSpec {
         var output = ""
         var thinkingMachine = ThinkingStateMachine()
         var pendingToolCalls: [(id: String, tool: ToolDefinition, args: [String: Any], argsJson: String)] = []
+        var rawTokenLog = ""
 
-        let chat = buildChatMessages(prompt: prompt, toolResults: toolResults, depth: depth)
-        let userInput = UserInput(
-            chat: chat,
-            tools: !self.toolSchemas.isEmpty ? self.toolSchemas : nil
+        let specialTokenPattern = try? NSRegularExpression(
+            pattern: "<\\|(?:im_end|im_start|endoftext|end|pad)\\|>",
+            options: []
         )
 
-        let lmInput = try await container.prepare(input: userInput)
+        log("perform_gen_events depth=\(depth) prompt=\(prompt.count)chars toolResults=\(toolResults?.count ?? 0)")
+
+        let chat = buildChatMessages(prompt: prompt, toolResults: toolResults, depth: depth)
+        let lmInput = try await prepareInput(container: container, chat: chat)
+        log("perform_gen_events input_prepared messages=\(chat.count) maxTokens=\(self.maxTokens) temperature=\(self.temperature)")
+
+        if self.lastInputContainedThinkTag {
+            let seed = thinkingMachine.process(token: "<think>")
+            for seedOutput in seed {
+                switch seedOutput {
+                case .thinkingStart:
+                    log("thinking_seeded_events")
+                    emitter.emitThinkingStart()
+                default:
+                    break
+                }
+            }
+        }
 
         let stream = try await container.perform { context in
-            let parameters = GenerateParameters(maxTokens: 2048, temperature: 0.7)
+            let parameters = GenerateParameters(maxTokens: self.maxTokens, temperature: Float(self.temperature))
             return try MLXLMCommon.generate(
                 input: lmInput,
                 parameters: parameters,
@@ -456,27 +562,53 @@ class HybridLLM: HybridLLMSpec {
             )
         }
 
+        var chunkCount = 0
         for await generation in stream {
-            if Task.isCancelled { break }
+            if Task.isCancelled {
+                log("perform_gen_events cancelled at chunk=\(chunkCount)")
+                break
+            }
 
             switch generation {
             case .chunk(let text):
+                chunkCount += 1
+                rawTokenLog += text
+                if chunkCount <= 20 || chunkCount % 50 == 0 {
+                    log("raw_chunk_events[\(chunkCount)] \(text.debugDescription)")
+                }
+
                 let outputs = thinkingMachine.process(token: text)
 
                 for machineOutput in outputs {
                     switch machineOutput {
                     case .token(let token):
-                        output += token
-                        emitter.emitToken(token)
-                        onTokenProcessed()
+                        var cleaned = token
+                        if let regex = specialTokenPattern {
+                            let before = cleaned
+                            cleaned = regex.stringByReplacingMatches(
+                                in: cleaned,
+                                range: NSRange(cleaned.startIndex..., in: cleaned),
+                                withTemplate: ""
+                            )
+                            if before != cleaned {
+                                log("stripped_special_events: \(before.debugDescription) -> \(cleaned.debugDescription)")
+                            }
+                        }
+                        if !cleaned.isEmpty {
+                            output += cleaned
+                            emitter.emitToken(cleaned)
+                            onTokenProcessed()
+                        }
 
                     case .thinkingStart:
+                        log("thinking_start_events at chunk=\(chunkCount)")
                         emitter.emitThinkingStart()
 
                     case .thinkingChunk(let chunk):
                         emitter.emitThinkingChunk(chunk)
 
                     case .thinkingEnd(let content):
+                        log("thinking_end_events at chunk=\(chunkCount)")
                         emitter.emitThinkingEnd(content)
                     }
                 }
@@ -497,12 +629,15 @@ class HybridLLM: HybridLLMSpec {
                 pendingToolCalls.append((id: toolCallId, tool: tool, args: argsDict, argsJson: argsJson))
 
             case .info(let info):
-                log("Generation info: \(info.generationTokenCount) tokens, \(String(format: "%.1f", info.tokensPerSecond)) tokens/s")
+                log("gen_info_events chunks=\(chunkCount) genTokens=\(info.generationTokenCount) tps=\(String(format: "%.1f", info.tokensPerSecond))")
                 let generationTime = info.tokensPerSecond > 0 ? Double(info.generationTokenCount) / info.tokensPerSecond * 1000 : 0
                 onGenerationInfo(info.generationTokenCount, generationTime)
             }
         }
 
+        log("perform_gen_events_loop_done chunks=\(chunkCount) output=\(output.count)chars")
+        log("raw_output_events: \(rawTokenLog)")
+
         let flushOutputs = thinkingMachine.flush()
         for machineOutput in flushOutputs {
             switch machineOutput {
@@ -592,18 +727,42 @@ class HybridLLM: HybridLLMSpec {
         }
 
         var output = ""
+        var thinkingMachine = ThinkingStateMachine()
         var pendingToolCalls: [(tool: ToolDefinition, args: [String: Any], argsJson: String)] = []
+        var rawTokenLog = ""
 
-        let chat = buildChatMessages(prompt: prompt, toolResults: toolResults, depth: depth)
-        let userInput = UserInput(
-            chat: chat,
-            tools: !self.toolSchemas.isEmpty ? self.toolSchemas : nil
+        let specialTokenPattern = try? NSRegularExpression(
+            pattern: "<\\|(?:im_end|im_start|endoftext|end|pad)\\|>",
+            options: []
         )
 
-        let lmInput = try await container.prepare(input: userInput)
+        log("perform_gen depth=\(depth) prompt=\(prompt.count)chars toolResults=\(toolResults?.count ?? 0)")
+
+        let chat = buildChatMessages(prompt: prompt, toolResults: toolResults, depth: depth)
+        let lmInput = try await prepareInput(container: container, chat: chat)
+        log("perform_gen input_prepared messages=\(chat.count) maxTokens=\(self.maxTokens) temperature=\(self.temperature)")
+
+        /*
+         When the chat template injects <think> at the end of the prompt,
+         the model generates thinking content directly — the opening tag
+         is NOT part of the generated stream. Seed the state machine so
+         the TS layer receives <think> and sets isThinking = true.
+        */
+        if self.lastInputContainedThinkTag {
+            let seed = thinkingMachine.process(token: "<think>")
+            for seedOutput in seed {
+                switch seedOutput {
+                case .thinkingStart:
+                    log("thinking_seeded")
+                    onToken("<think>")
+                default:
+                    break
+                }
+            }
+        }
 
         let stream = try await container.perform { context in
-            let parameters = GenerateParameters(maxTokens: 2048, temperature: 0.7)
+            let parameters = GenerateParameters(maxTokens: self.maxTokens, temperature: Float(self.temperature))
             return try MLXLMCommon.generate(
                 input: lmInput,
                 parameters: parameters,
@@ -611,13 +770,55 @@ class HybridLLM: HybridLLMSpec {
             )
         }
 
+        var chunkCount = 0
         for await generation in stream {
-            if Task.isCancelled { break }
+            if Task.isCancelled {
+                log("perform_gen cancelled at chunk=\(chunkCount)")
+                break
+            }
 
             switch generation {
             case .chunk(let text):
-                output += text
-                onToken(text)
+                chunkCount += 1
+                rawTokenLog += text
+                if chunkCount <= 20 || chunkCount % 50 == 0 {
+                    log("raw_chunk[\(chunkCount)] \(text.debugDescription)")
+                }
+
+                let outputs = thinkingMachine.process(token: text)
+
+                for machineOutput in outputs {
+                    switch machineOutput {
+                    case .token(let token):
+                        var cleaned = token
+                        if let regex = specialTokenPattern {
+                            let before = cleaned
+                            cleaned = regex.stringByReplacingMatches(
+                                in: cleaned,
+                                range: NSRange(cleaned.startIndex..., in: cleaned),
+                                withTemplate: ""
+                            )
+                            if before != cleaned {
+                                log("stripped_special: \(before.debugDescription) -> \(cleaned.debugDescription)")
+                            }
+                        }
+                        if !cleaned.isEmpty {
+                            output += cleaned
+                            onToken(cleaned)
+                        }
+
+                    case .thinkingStart:
+                        log("thinking_start at chunk=\(chunkCount)")
+                        onToken("<think>")
+
+                    case .thinkingChunk(let chunk):
+                        onToken(chunk)
+
+                    case .thinkingEnd:
+                        log("thinking_end at chunk=\(chunkCount)")
+                        onToken("</think>")
+                    }
+                }
 
             case .toolCall(let toolCall):
                 log("Tool call detected: \(toolCall.function.name)")
@@ -634,7 +835,38 @@ class HybridLLM: HybridLLMSpec {
                 onToolCall(toolCall.function.name, argsJson)
 
             case .info(let info):
-                log("Generation info: \(info.generationTokenCount) tokens, \(String(format: "%.1f", info.tokensPerSecond)) tokens/s")
+                log("gen_info chunks=\(chunkCount) genTokens=\(info.generationTokenCount) tps=\(String(format: "%.1f", info.tokensPerSecond))")
+            }
+        }
+
+        log("perform_gen_loop_done chunks=\(chunkCount) output=\(output.count)chars")
+        log("raw_output: \(rawTokenLog)")
+
+        let flushOutputs = thinkingMachine.flush()
+        if !flushOutputs.isEmpty {
+            log("flush_outputs count=\(flushOutputs.count)")
+        }
+        for machineOutput in flushOutputs {
+            switch machineOutput {
+            case .token(let token):
+                var cleaned = token
+                if let regex = specialTokenPattern {
+                    cleaned = regex.stringByReplacingMatches(
+                        in: cleaned,
+                        range: NSRange(cleaned.startIndex..., in: cleaned),
+                        withTemplate: ""
+                    )
+                }
+                if !cleaned.isEmpty {
+                    output += cleaned
+                    onToken(cleaned)
+                }
+            case .thinkingStart:
+                onToken("<think>")
+            case .thinkingChunk(let chunk):
+                onToken(chunk)
+            case .thinkingEnd:
+                onToken("</think>")
             }
         }
 
@@ -688,6 +920,7 @@ class HybridLLM: HybridLLMSpec {
             return output + continuation
         }
 
+        log("perform_gen_result output=\(output.count)chars result: \(output)")
         return output
     }
 
@@ -771,7 +1004,14 @@ class HybridLLM: HybridLLMSpec {
     }
 
     func clearHistory() throws {
+        log("clear_history before=\(messageHistory.count) messages")
+        for (i, msg) in messageHistory.enumerated() {
+            log("  clearing[\(i):\(msg.role)] \(msg.content)")
+        }
         messageHistory = []
-        log("Message history cleared")
+        if let container = self.container {
+            self.session = ChatSession(container, instructions: self.systemPrompt)
+        }
+        log("clear_history done session_reset")
     }
 }

package/lib/module/llm.js

@@ -188,6 +188,24 @@ export const LLM = {
   },
   set systemPrompt(value) {
     getInstance().systemPrompt = value;
+  },
+  get maxTokens() {
+    return getInstance().maxTokens;
+  },
+  set maxTokens(value) {
+    getInstance().maxTokens = value;
+  },
+  get temperature() {
+    return getInstance().temperature;
+  },
+  set temperature(value) {
+    getInstance().temperature = value;
+  },
+  get enableThinking() {
+    return getInstance().enableThinking;
+  },
+  set enableThinking(value) {
+    getInstance().enableThinking = value;
   }
 };
 //# sourceMappingURL=llm.js.map

package/lib/module/llm.js.map

@@ -1 +1 @@
-{"version":3,"names":["NitroModules","instance","getInstance","createHybridObject","LLM","load","modelId","options","generate","prompt","stream","onToken","onToolCall","accumulatedToolCalls","name","argsJson","args","JSON","parse","toolCall","arguments","push","allToolCalls","streamWithEvents","onEvent","eventJson","event","stop","unload","getLastGenerationStats","getHistory","clearHistory","isLoaded","isGenerating","debug","value","systemPrompt"],"sourceRoot":"../../src","sources":["llm.ts"],"mappings":";;AAAA,SAASA,YAAY,QAAQ,4BAA4B;AAUzD,IAAIC,QAAwB,GAAG,IAAI;AAiBnC,SAASC,WAAWA,CAAA,EAAY;EAC9B,IAAI,CAACD,QAAQ,EAAE;IACbA,QAAQ,GAAGD,YAAY,CAACG,kBAAkB,CAAU,KAAK,CAAC;EAC5D;EACA,OAAOF,QAAQ;AACjB;;AAEA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA,OAAO,MAAMG,GAAG,GAAG;EACjB;AACF;AACA;AACA;AACA;EACEC,IAAIA,CAACC,OAAe,EAAEC,OAAuB,EAAiB;IAC5D,OAAOL,WAAW,CAAC,CAAC,CAACG,IAAI,CAACC,OAAO,EAAEC,OAAO,CAAC;EAC7C,CAAC;EAED;AACF;AACA;AACA;AACA;AACA;EACEC,QAAQA,CAACC,MAAc,EAAmB;IACxC,OAAOP,WAAW,CAAC,CAAC,CAACM,QAAQ,CAACC,MAAM,CAAC;EACvC,CAAC;EAED;AACF;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;EACEC,MAAMA,CACJD,MAAc,EACdE,OAAgC,EAChCC,UAA6C,EAC5B;IACjB,MAAMC,oBAAoC,GAAG,EAAE;IAE/C,OAAOX,WAAW,CAAC,CAAC,CAACQ,MAAM,CAACD,MAAM,EAAEE,OAAO,EAAE,CAACG,IAAY,EAAEC,QAAgB,KAAK;MAC/E,IAAIH,UAAU,EAAE;QACd,IAAI;UACF,MAAMI,IAAI,GAAGC,IAAI,CAACC,KAAK,CAACH,QAAQ,CAA4B;UAC5D,MAAMI,QAAQ,GAAG;YAAEL,IAAI;YAAEM,SAAS,EAAEJ;UAAK,CAAC;UAC1CH,oBAAoB,CAACQ,IAAI,CAACF,QAAQ,CAAC;UACnCP,UAAU,CAAC;YACTO,QAAQ;YACRG,YAAY,EAAE,CAAC,GAAGT,oBAAoB;UACxC,CAAC,CAAC;QACJ,CAAC,CAAC,MAAM;UACN,MAAMM,QAAQ,GAAG;YAAEL,IAAI;YAAEM,SAAS,EAAE,CAAC;UAAE,CAAC;UACxCP,oBAAoB,CAACQ,IAAI,CAACF,QAAQ,CAAC;UACnCP,UAAU,CAAC;YACTO,QAAQ;YACRG,YAAY,EAAE,CAAC,GAAGT,oBAAoB;UACxC,CAAC,CAAC;QACJ;MACF;IACF,CAAC,CAAC;EACJ,CAAC;EAED;AACF;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;EACEU,gBAAgBA,CAACd,MAAc,EAAEe,OAAsB,EAAmB;IACxE,OAAOtB,WAAW,CAAC,CAAC,CAACqB,gBAAgB,CAACd,MAAM,EAAGgB,SAAiB,IAAK;MACnE,IAAI;QACF,MAAMC,KAAK,GAAGT,IAAI,CAACC,KAAK,CAACO,SAAS,CAAgB;QAClDD,OAAO,CAACE,KAAK,CAAC;MAChB,CAAC,CAAC,MAAM;QACN;MAAA;IAEJ,CAAC,CAAC;EACJ,CAAC;EAED;AACF;AACA;EACEC,IAAIA,CAAA,EAAS;IACXzB,WAAW,CAAC,CAAC,CAACyB,IAAI,CAAC,CAAC;EACtB,CAAC;EAED;AACF;AACA;AACA;EACEC,MAAMA,CAAA,EAAS;IACb1B,WAAW,CAAC,CAAC,CAAC0B,MAAM,CAAC,CAAC;EACxB,CAAC;EAED;AACF;AACA;AACA;EACEC,sBAAsBA,CAAA,EAAoB;IACxC,OAAO3B,WAAW,CAAC,CAAC,CAAC2B,sBAAsB,CAAC,CAAC;EAC/C,CAAC;EAED;AACF;AACA;AACA;EACEC,UAAUA,CAAA,EAAc;IACtB,OAAO5B,WAAW,CAAC,CAAC,CAAC4B,UAAU,CAAC,CAAC;EACnC,CAAC;EAED;AACF;AACA;EACEC,YAAYA,CAAA,EAAS;IACnB7B,WAAW,CAAC,CAAC,CAAC6B,YAAY,CAAC,CAAC;EAC9B,CAAC;EAED;EACA,IAAIC,QAAQA,CAAA,EAAY;IACtB,OAAO9B,WAAW,CAAC,CAAC,CAAC8B,QAAQ;EAC/B,CAAC;EAED;EACA,IAAIC,YAAYA,CAAA,EAAY;IAC1B,OAAO/B,WAAW,CAAC,CAAC,CAAC+B,YAAY;EACnC,CAAC;EAED;EACA,IAAI3B,OAAOA,CAAA,EAAW;IACpB,OAAOJ,WAAW,CAAC,CAAC,CAACI,OAAO;EAC9B,CAAC;EAED;EACA,IAAI4B,KAAKA,CAAA,EAAY;IACnB,OAAOhC,WAAW,CAAC,CAAC,CAACgC,KAAK;EAC5B,CAAC;EAED,IAAIA,KAAKA,CAACC,KAAc,EAAE;IACxBjC,WAAW,CAAC,CAAC,CAACgC,KAAK,GAAGC,KAAK;EAC7B,CAAC;EAED;AACF;AACA;AACA;AACA;EACE,IAAIC,YAAYA,CAAA,EAAW;IACzB,OAAOlC,WAAW,CAAC,CAAC,CAACkC,YAAY;EACnC,CAAC;EAED,IAAIA,YAAYA,CAACD,KAAa,EAAE;IAC9BjC,WAAW,CAAC,CAAC,CAACkC,YAAY,GAAGD,KAAK;EACpC;AACF,CAAC","ignoreList":[]}
+{"version":3,"names":["NitroModules","instance","getInstance","createHybridObject","LLM","load","modelId","options","generate","prompt","stream","onToken","onToolCall","accumulatedToolCalls","name","argsJson","args","JSON","parse","toolCall","arguments","push","allToolCalls","streamWithEvents","onEvent","eventJson","event","stop","unload","getLastGenerationStats","getHistory","clearHistory","isLoaded","isGenerating","debug","value","systemPrompt","maxTokens","temperature","enableThinking"],"sourceRoot":"../../src","sources":["llm.ts"],"mappings":";;AAAA,SAASA,YAAY,QAAQ,4BAA4B;AAUzD,IAAIC,QAAwB,GAAG,IAAI;AAiBnC,SAASC,WAAWA,CAAA,EAAY;EAC9B,IAAI,CAACD,QAAQ,EAAE;IACbA,QAAQ,GAAGD,YAAY,CAACG,kBAAkB,CAAU,KAAK,CAAC;EAC5D;EACA,OAAOF,QAAQ;AACjB;;AAEA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA,OAAO,MAAMG,GAAG,GAAG;EACjB;AACF;AACA;AACA;AACA;EACEC,IAAIA,CAACC,OAAe,EAAEC,OAAuB,EAAiB;IAC5D,OAAOL,WAAW,CAAC,CAAC,CAACG,IAAI,CAACC,OAAO,EAAEC,OAAO,CAAC;EAC7C,CAAC;EAED;AACF;AACA;AACA;AACA;AACA;EACEC,QAAQA,CAACC,MAAc,EAAmB;IACxC,OAAOP,WAAW,CAAC,CAAC,CAACM,QAAQ,CAACC,MAAM,CAAC;EACvC,CAAC;EAED;AACF;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;EACEC,MAAMA,CACJD,MAAc,EACdE,OAAgC,EAChCC,UAA6C,EAC5B;IACjB,MAAMC,oBAAoC,GAAG,EAAE;IAE/C,OAAOX,WAAW,CAAC,CAAC,CAACQ,MAAM,CAACD,MAAM,EAAEE,OAAO,EAAE,CAACG,IAAY,EAAEC,QAAgB,KAAK;MAC/E,IAAIH,UAAU,EAAE;QACd,IAAI;UACF,MAAMI,IAAI,GAAGC,IAAI,CAACC,KAAK,CAACH,QAAQ,CAA4B;UAC5D,MAAMI,QAAQ,GAAG;YAAEL,IAAI;YAAEM,SAAS,EAAEJ;UAAK,CAAC;UAC1CH,oBAAoB,CAACQ,IAAI,CAACF,QAAQ,CAAC;UACnCP,UAAU,CAAC;YACTO,QAAQ;YACRG,YAAY,EAAE,CAAC,GAAGT,oBAAoB;UACxC,CAAC,CAAC;QACJ,CAAC,CAAC,MAAM;UACN,MAAMM,QAAQ,GAAG;YAAEL,IAAI;YAAEM,SAAS,EAAE,CAAC;UAAE,CAAC;UACxCP,oBAAoB,CAACQ,IAAI,CAACF,QAAQ,CAAC;UACnCP,UAAU,CAAC;YACTO,QAAQ;YACRG,YAAY,EAAE,CAAC,GAAGT,oBAAoB;UACxC,CAAC,CAAC;QACJ;MACF;IACF,CAAC,CAAC;EACJ,CAAC;EAED;AACF;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;EACEU,gBAAgBA,CAACd,MAAc,EAAEe,OAAsB,EAAmB;IACxE,OAAOtB,WAAW,CAAC,CAAC,CAACqB,gBAAgB,CAACd,MAAM,EAAGgB,SAAiB,IAAK;MACnE,IAAI;QACF,MAAMC,KAAK,GAAGT,IAAI,CAACC,KAAK,CAACO,SAAS,CAAgB;QAClDD,OAAO,CAACE,KAAK,CAAC;MAChB,CAAC,CAAC,MAAM;QACN;MAAA;IAEJ,CAAC,CAAC;EACJ,CAAC;EAED;AACF;AACA;EACEC,IAAIA,CAAA,EAAS;IACXzB,WAAW,CAAC,CAAC,CAACyB,IAAI,CAAC,CAAC;EACtB,CAAC;EAED;AACF;AACA;AACA;EACEC,MAAMA,CAAA,EAAS;IACb1B,WAAW,CAAC,CAAC,CAAC0B,MAAM,CAAC,CAAC;EACxB,CAAC;EAED;AACF;AACA;AACA;EACEC,sBAAsBA,CAAA,EAAoB;IACxC,OAAO3B,WAAW,CAAC,CAAC,CAAC2B,sBAAsB,CAAC,CAAC;EAC/C,CAAC;EAED;AACF;AACA;AACA;EACEC,UAAUA,CAAA,EAAc;IACtB,OAAO5B,WAAW,CAAC,CAAC,CAAC4B,UAAU,CAAC,CAAC;EACnC,CAAC;EAED;AACF;AACA;EACEC,YAAYA,CAAA,EAAS;IACnB7B,WAAW,CAAC,CAAC,CAAC6B,YAAY,CAAC,CAAC;EAC9B,CAAC;EAED;EACA,IAAIC,QAAQA,CAAA,EAAY;IACtB,OAAO9B,WAAW,CAAC,CAAC,CAAC8B,QAAQ;EAC/B,CAAC;EAED;EACA,IAAIC,YAAYA,CAAA,EAAY;IAC1B,OAAO/B,WAAW,CAAC,CAAC,CAAC+B,YAAY;EACnC,CAAC;EAED;EACA,IAAI3B,OAAOA,CAAA,EAAW;IACpB,OAAOJ,WAAW,CAAC,CAAC,CAACI,OAAO;EAC9B,CAAC;EAED;EACA,IAAI4B,KAAKA,CAAA,EAAY;IACnB,OAAOhC,WAAW,CAAC,CAAC,CAACgC,KAAK;EAC5B,CAAC;EAED,IAAIA,KAAKA,CAACC,KAAc,EAAE;IACxBjC,WAAW,CAAC,CAAC,CAACgC,KAAK,GAAGC,KAAK;EAC7B,CAAC;EAED;AACF;AACA;AACA;AACA;EACE,IAAIC,YAAYA,CAAA,EAAW;IACzB,OAAOlC,WAAW,CAAC,CAAC,CAACkC,YAAY;EACnC,CAAC;EAED,IAAIA,YAAYA,CAACD,KAAa,EAAE;IAC9BjC,WAAW,CAAC,CAAC,CAACkC,YAAY,GAAGD,KAAK;EACpC,CAAC;EAED,IAAIE,SAASA,CAAA,EAAW;IACtB,OAAOnC,WAAW,CAAC,CAAC,CAACmC,SAAS;EAChC,CAAC;EAED,IAAIA,SAASA,CAACF,KAAa,EAAE;IAC3BjC,WAAW,CAAC,CAAC,CAACmC,SAAS,GAAGF,KAAK;EACjC,CAAC;EAED,IAAIG,WAAWA,CAAA,EAAW;IACxB,OAAOpC,WAAW,CAAC,CAAC,CAACoC,WAAW;EAClC,CAAC;EAED,IAAIA,WAAWA,CAACH,KAAa,EAAE;IAC7BjC,WAAW,CAAC,CAAC,CAACoC,WAAW,GAAGH,KAAK;EACnC,CAAC;EAED,IAAII,cAAcA,CAAA,EAAY;IAC5B,OAAOrC,WAAW,CAAC,CAAC,CAACqC,cAAc;EACrC,CAAC;EAED,IAAIA,cAAcA,CAACJ,KAAc,EAAE;IACjCjC,WAAW,CAAC,CAAC,CAACqC,cAAc,GAAGJ,KAAK;EACtC;AACF,CAAC","ignoreList":[]}

package/lib/typescript/src/llm.d.ts

@@ -125,5 +125,8 @@ export declare const LLM: {
      * @default "You are a helpful assistant."
      */
     systemPrompt: string;
+    maxTokens: number;
+    temperature: number;
+    enableThinking: boolean;
 };
 //# sourceMappingURL=llm.d.ts.map

package/lib/typescript/src/llm.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"llm.d.ts","sourceRoot":"","sources":["../../../src/llm.ts"],"names":[],"mappings":"AACA,OAAO,KAAK,EACV,eAAe,EACf,cAAc,EAEd,WAAW,EACZ,MAAM,mBAAmB,CAAA;AAE1B,MAAM,MAAM,aAAa,GAAG,CAAC,KAAK,EAAE,WAAW,KAAK,IAAI,CAAA;AAIxD,MAAM,MAAM,OAAO,GAAG;IACpB,IAAI,EAAE,MAAM,GAAG,WAAW,GAAG,QAAQ,CAAA;IACrC,OAAO,EAAE,MAAM,CAAA;CAChB,CAAA;AAED,MAAM,MAAM,YAAY,GAAG;IACzB,IAAI,EAAE,MAAM,CAAA;IACZ,SAAS,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAA;CACnC,CAAA;AAED,MAAM,MAAM,cAAc,GAAG;IAC3B,QAAQ,EAAE,YAAY,CAAA;IACtB,YAAY,EAAE,YAAY,EAAE,CAAA;CAC7B,CAAA;AASD;;;;;;;;;;;;;;;;;;;;;GAqBG;AACH,eAAO,MAAM,GAAG;IACd;;;;OAIG;kBACW,MAAM,WAAW,cAAc,GAAG,OAAO,CAAC,IAAI,CAAC;IAI7D;;;;;OAKG;qBACc,MAAM,GAAG,OAAO,CAAC,MAAM,CAAC;IAIzC;;;;;;;;;OASG;mBAEO,MAAM,WACL,CAAC,KAAK,EAAE,MAAM,KAAK,IAAI,eACnB,CAAC,MAAM,EAAE,cAAc,KAAK,IAAI,GAC5C,OAAO,CAAC,MAAM,CAAC;IAyBlB;;;;;;;;;;;;;;;;;;;;;;;;;;;OA2BG;6BACsB,MAAM,WAAW,aAAa,GAAG,OAAO,CAAC,MAAM,CAAC;IAWzE;;OAEG;YACK,IAAI;IAIZ;;;OAGG;cACO,IAAI;IAId;;;OAGG;8BACuB,eAAe;IAIzC;;;OAGG;kBACW,OAAO,EAAE;IAIvB;;OAEG;oBACa,IAAI;IAIpB,mEAAmE;uBACnD,OAAO;IAIvB,gDAAgD;2BAC5B,OAAO;IAI3B,oEAAoE;sBACrD,MAAM;IAIrB,sCAAsC;WACzB,OAAO;IAQpB;;;;OAIG;kBACiB,MAAM;CAO3B,CAAA"}
+{"version":3,"file":"llm.d.ts","sourceRoot":"","sources":["../../../src/llm.ts"],"names":[],"mappings":"AACA,OAAO,KAAK,EACV,eAAe,EACf,cAAc,EAEd,WAAW,EACZ,MAAM,mBAAmB,CAAA;AAE1B,MAAM,MAAM,aAAa,GAAG,CAAC,KAAK,EAAE,WAAW,KAAK,IAAI,CAAA;AAIxD,MAAM,MAAM,OAAO,GAAG;IACpB,IAAI,EAAE,MAAM,GAAG,WAAW,GAAG,QAAQ,CAAA;IACrC,OAAO,EAAE,MAAM,CAAA;CAChB,CAAA;AAED,MAAM,MAAM,YAAY,GAAG;IACzB,IAAI,EAAE,MAAM,CAAA;IACZ,SAAS,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAA;CACnC,CAAA;AAED,MAAM,MAAM,cAAc,GAAG;IAC3B,QAAQ,EAAE,YAAY,CAAA;IACtB,YAAY,EAAE,YAAY,EAAE,CAAA;CAC7B,CAAA;AASD;;;;;;;;;;;;;;;;;;;;;GAqBG;AACH,eAAO,MAAM,GAAG;IACd;;;;OAIG;kBACW,MAAM,WAAW,cAAc,GAAG,OAAO,CAAC,IAAI,CAAC;IAI7D;;;;;OAKG;qBACc,MAAM,GAAG,OAAO,CAAC,MAAM,CAAC;IAIzC;;;;;;;;;OASG;mBAEO,MAAM,WACL,CAAC,KAAK,EAAE,MAAM,KAAK,IAAI,eACnB,CAAC,MAAM,EAAE,cAAc,KAAK,IAAI,GAC5C,OAAO,CAAC,MAAM,CAAC;IAyBlB;;;;;;;;;;;;;;;;;;;;;;;;;;;OA2BG;6BACsB,MAAM,WAAW,aAAa,GAAG,OAAO,CAAC,MAAM,CAAC;IAWzE;;OAEG;YACK,IAAI;IAIZ;;;OAGG;cACO,IAAI;IAId;;;OAGG;8BACuB,eAAe;IAIzC;;;OAGG;kBACW,OAAO,EAAE;IAIvB;;OAEG;oBACa,IAAI;IAIpB,mEAAmE;uBACnD,OAAO;IAIvB,gDAAgD;2BAC5B,OAAO;IAI3B,oEAAoE;sBACrD,MAAM;IAIrB,sCAAsC;WACzB,OAAO;IAQpB;;;;OAIG;kBACiB,MAAM;eAQT,MAAM;iBAQJ,MAAM;oBAQH,OAAO;CAO9B,CAAA"}

package/lib/typescript/src/specs/LLM.nitro.d.ts

@@ -151,6 +151,12 @@ export interface LLM extends HybridObject<{
     debug: boolean;
     /** System prompt used when loading the model */
     systemPrompt: string;
+    /** Maximum number of tokens to generate */
+    maxTokens: number;
+    /** Sampling temperature (0 = deterministic, higher = more random) */
+    temperature: number;
+    /** Enable thinking mode for models that support it */
+    enableThinking: boolean;
 }
 /**
  * Supported parameter types for tool definitions.

package/lib/typescript/src/specs/LLM.nitro.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"LLM.nitro.d.ts","sourceRoot":"","sources":["../../../../src/specs/LLM.nitro.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,MAAM,EAAE,YAAY,EAAE,MAAM,4BAA4B,CAAA;AAEtE;;GAEG;AACH,MAAM,WAAW,eAAe;IAC9B,UAAU,EAAE,MAAM,CAAA;IAClB,eAAe,EAAE,MAAM,CAAA;IACvB,gBAAgB,EAAE,MAAM,CAAA;IACxB,SAAS,EAAE,MAAM,CAAA;IACjB,iBAAiB,EAAE,MAAM,CAAA;CAC1B;AAED,MAAM,WAAW,oBAAoB;IACnC,IAAI,EAAE,kBAAkB,CAAA;IACxB,SAAS,EAAE,MAAM,CAAA;CAClB;AAED,MAAM,WAAW,UAAU;IACzB,IAAI,EAAE,OAAO,CAAA;IACb,KAAK,EAAE,MAAM,CAAA;CACd;AAED,MAAM,WAAW,kBAAkB;IACjC,IAAI,EAAE,gBAAgB,CAAA;IACtB,SAAS,EAAE,MAAM,CAAA;CAClB;AAED,MAAM,WAAW,kBAAkB;IACjC,IAAI,EAAE,gBAAgB,CAAA;IACtB,KAAK,EAAE,MAAM,CAAA;CACd;AAED,MAAM,WAAW,gBAAgB;IAC/B,IAAI,EAAE,cAAc,CAAA;IACpB,OAAO,EAAE,MAAM,CAAA;IACf,SAAS,EAAE,MAAM,CAAA;CAClB;AAED,MAAM,WAAW,kBAAkB;IACjC,IAAI,EAAE,iBAAiB,CAAA;IACvB,EAAE,EAAE,MAAM,CAAA;IACV,IAAI,EAAE,MAAM,CAAA;IACZ,SAAS,EAAE,MAAM,CAAA;CAClB;AAED,MAAM,WAAW,sBAAsB;IACrC,IAAI,EAAE,qBAAqB,CAAA;IAC3B,EAAE,EAAE,MAAM,CAAA;CACX;AAED,MAAM,WAAW,sBAAsB;IACrC,IAAI,EAAE,qBAAqB,CAAA;IAC3B,EAAE,EAAE,MAAM,CAAA;IACV,MAAM,EAAE,MAAM,CAAA;CACf;AAED,MAAM,WAAW,mBAAmB;IAClC,IAAI,EAAE,kBAAkB,CAAA;IACxB,EAAE,EAAE,MAAM,CAAA;IACV,KAAK,EAAE,MAAM,CAAA;CACd;AAED,MAAM,WAAW,kBAAkB;IACjC,IAAI,EAAE,gBAAgB,CAAA;IACtB,OAAO,EAAE,MAAM,CAAA;IACf,KAAK,EAAE,eAAe,CAAA;CACvB;AAED,MAAM,MAAM,WAAW,GACnB,oBAAoB,GACpB,UAAU,GACV,kBAAkB,GAClB,kBAAkB,GAClB,gBAAgB,GAChB,kBAAkB,GAClB,sBAAsB,GACtB,sBAAsB,GACtB,mBAAmB,GACnB,kBAAkB,CAAA;AAEtB,MAAM,WAAW,UAAU;IACzB,IAAI,EAAE,MAAM,CAAA;IACZ,OAAO,EAAE,MAAM,CAAA;CAChB;AAED;;GAEG;AACH,MAAM,WAAW,aAAa;IAC5B,IAAI,EAAE,MAAM,CAAA;IACZ,IAAI,EAAE,MAAM,CAAA;IACZ,WAAW,EAAE,MAAM,CAAA;IACnB,QAAQ,EAAE,OAAO,CAAA;CAClB;AAED;;GAEG;AACH,MAAM,WAAW,cAAc;IAC7B,IAAI,EAAE,MAAM,CAAA;IACZ,WAAW,EAAE,MAAM,CAAA;IACnB,UAAU,EAAE,aAAa,EAAE,CAAA;IAC3B,OAAO,EAAE,CAAC,IAAI,EAAE,MAAM,KAAK,OAAO,CAAC,MAAM,CAAC,CAAA;CAC3C;AAED;GACG;AACH,MAAM,WAAW,cAAc;IAC7B,mDAAmD;IACnD,UAAU,CAAC,EAAE,CAAC,QAAQ,EAAE,MAAM,KAAK,IAAI,CAAA;IACvC,iDAAiD;IACjD,iBAAiB,CAAC,EAAE,UAAU,EAAE,CAAA;IAChC,sDAAsD;IACtD,aAAa,CAAC,EAAE,OAAO,CAAA;IACvB,4CAA4C;IAC5C,KAAK,CAAC,EAAE,cAAc,EAAE,CAAA;CACzB;AAED;;;GAGG;AACH,MAAM,WAAW,GAAI,SAAQ,YAAY,CAAC;IAAE,GAAG,EAAE,OAAO,CAAA;CAAE,CAAC;IACzD;;;;OAIG;IACH,IAAI,CAAC,OAAO,EAAE,MAAM,EAAE,OAAO,CAAC,EAAE,cAAc,GAAG,OAAO,CAAC,IAAI,CAAC,CAAA;IAE9D;;;;OAIG;IACH,QAAQ,CAAC,MAAM,EAAE,MAAM,GAAG,OAAO,CAAC,MAAM,CAAC,CAAA;IAEzC;;;;;;;OAOG;IACH,MAAM,CACJ,MAAM,EAAE,MAAM,EACd,OAAO,EAAE,CAAC,KAAK,EAAE,MAAM,KAAK,IAAI,EAChC,UAAU,CAAC,EAAE,CAAC,QAAQ,EAAE,MAAM,EAAE,IAAI,EAAE,MAAM,KAAK,IAAI,GACpD,OAAO,CAAC,MAAM,CAAC,CAAA;IAElB,gBAAgB,CACd,MAAM,EAAE,MAAM,EACd,OAAO,EAAE,CAAC,SAAS,EAAE,MAAM,KAAK,IAAI,GACnC,OAAO,CAAC,MAAM,CAAC,CAAA;IAElB;;OAEG;IACH,IAAI,IAAI,IAAI,CAAA;IAEZ;;OAEG;IACH,MAAM,IAAI,IAAI,CAAA;IAEd;;;OAGG;IACH,sBAAsB,IAAI,eAAe,CAAA;IAEzC;;;OAGG;IACH,UAAU,IAAI,UAAU,EAAE,CAAA;IAE1B;;OAEG;IACH,YAAY,IAAI,IAAI,CAAA;IAEpB,0CAA0C;IAC1C,QAAQ,CAAC,QAAQ,EAAE,OAAO,CAAA;IAC1B,gDAAgD;IAChD,QAAQ,CAAC,YAAY,EAAE,OAAO,CAAA;IAC9B,2CAA2C;IAC3C,QAAQ,CAAC,OAAO,EAAE,MAAM,CAAA;IAExB,2BAA2B;IAC3B,KAAK,EAAE,OAAO,CAAA;IACd,gDAAgD;IAChD,YAAY,EAAE,MAAM,CAAA;CACrB;AAED;;;GAGG;AACH,MAAM,MAAM,iBAAiB,GAAG,QAAQ,GAAG,QAAQ,GAAG,SAAS,GAAG,OAAO,GAAG,QAAQ,CAAA"}
+{"version":3,"file":"LLM.nitro.d.ts","sourceRoot":"","sources":["../../../../src/specs/LLM.nitro.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,MAAM,EAAE,YAAY,EAAE,MAAM,4BAA4B,CAAA;AAEtE;;GAEG;AACH,MAAM,WAAW,eAAe;IAC9B,UAAU,EAAE,MAAM,CAAA;IAClB,eAAe,EAAE,MAAM,CAAA;IACvB,gBAAgB,EAAE,MAAM,CAAA;IACxB,SAAS,EAAE,MAAM,CAAA;IACjB,iBAAiB,EAAE,MAAM,CAAA;CAC1B;AAED,MAAM,WAAW,oBAAoB;IACnC,IAAI,EAAE,kBAAkB,CAAA;IACxB,SAAS,EAAE,MAAM,CAAA;CAClB;AAED,MAAM,WAAW,UAAU;IACzB,IAAI,EAAE,OAAO,CAAA;IACb,KAAK,EAAE,MAAM,CAAA;CACd;AAED,MAAM,WAAW,kBAAkB;IACjC,IAAI,EAAE,gBAAgB,CAAA;IACtB,SAAS,EAAE,MAAM,CAAA;CAClB;AAED,MAAM,WAAW,kBAAkB;IACjC,IAAI,EAAE,gBAAgB,CAAA;IACtB,KAAK,EAAE,MAAM,CAAA;CACd;AAED,MAAM,WAAW,gBAAgB;IAC/B,IAAI,EAAE,cAAc,CAAA;IACpB,OAAO,EAAE,MAAM,CAAA;IACf,SAAS,EAAE,MAAM,CAAA;CAClB;AAED,MAAM,WAAW,kBAAkB;IACjC,IAAI,EAAE,iBAAiB,CAAA;IACvB,EAAE,EAAE,MAAM,CAAA;IACV,IAAI,EAAE,MAAM,CAAA;IACZ,SAAS,EAAE,MAAM,CAAA;CAClB;AAED,MAAM,WAAW,sBAAsB;IACrC,IAAI,EAAE,qBAAqB,CAAA;IAC3B,EAAE,EAAE,MAAM,CAAA;CACX;AAED,MAAM,WAAW,sBAAsB;IACrC,IAAI,EAAE,qBAAqB,CAAA;IAC3B,EAAE,EAAE,MAAM,CAAA;IACV,MAAM,EAAE,MAAM,CAAA;CACf;AAED,MAAM,WAAW,mBAAmB;IAClC,IAAI,EAAE,kBAAkB,CAAA;IACxB,EAAE,EAAE,MAAM,CAAA;IACV,KAAK,EAAE,MAAM,CAAA;CACd;AAED,MAAM,WAAW,kBAAkB;IACjC,IAAI,EAAE,gBAAgB,CAAA;IACtB,OAAO,EAAE,MAAM,CAAA;IACf,KAAK,EAAE,eAAe,CAAA;CACvB;AAED,MAAM,MAAM,WAAW,GACnB,oBAAoB,GACpB,UAAU,GACV,kBAAkB,GAClB,kBAAkB,GAClB,gBAAgB,GAChB,kBAAkB,GAClB,sBAAsB,GACtB,sBAAsB,GACtB,mBAAmB,GACnB,kBAAkB,CAAA;AAEtB,MAAM,WAAW,UAAU;IACzB,IAAI,EAAE,MAAM,CAAA;IACZ,OAAO,EAAE,MAAM,CAAA;CAChB;AAED;;GAEG;AACH,MAAM,WAAW,aAAa;IAC5B,IAAI,EAAE,MAAM,CAAA;IACZ,IAAI,EAAE,MAAM,CAAA;IACZ,WAAW,EAAE,MAAM,CAAA;IACnB,QAAQ,EAAE,OAAO,CAAA;CAClB;AAED;;GAEG;AACH,MAAM,WAAW,cAAc;IAC7B,IAAI,EAAE,MAAM,CAAA;IACZ,WAAW,EAAE,MAAM,CAAA;IACnB,UAAU,EAAE,aAAa,EAAE,CAAA;IAC3B,OAAO,EAAE,CAAC,IAAI,EAAE,MAAM,KAAK,OAAO,CAAC,MAAM,CAAC,CAAA;CAC3C;AAED;GACG;AACH,MAAM,WAAW,cAAc;IAC7B,mDAAmD;IACnD,UAAU,CAAC,EAAE,CAAC,QAAQ,EAAE,MAAM,KAAK,IAAI,CAAA;IACvC,iDAAiD;IACjD,iBAAiB,CAAC,EAAE,UAAU,EAAE,CAAA;IAChC,sDAAsD;IACtD,aAAa,CAAC,EAAE,OAAO,CAAA;IACvB,4CAA4C;IAC5C,KAAK,CAAC,EAAE,cAAc,EAAE,CAAA;CACzB;AAED;;;GAGG;AACH,MAAM,WAAW,GAAI,SAAQ,YAAY,CAAC;IAAE,GAAG,EAAE,OAAO,CAAA;CAAE,CAAC;IACzD;;;;OAIG;IACH,IAAI,CAAC,OAAO,EAAE,MAAM,EAAE,OAAO,CAAC,EAAE,cAAc,GAAG,OAAO,CAAC,IAAI,CAAC,CAAA;IAE9D;;;;OAIG;IACH,QAAQ,CAAC,MAAM,EAAE,MAAM,GAAG,OAAO,CAAC,MAAM,CAAC,CAAA;IAEzC;;;;;;;OAOG;IACH,MAAM,CACJ,MAAM,EAAE,MAAM,EACd,OAAO,EAAE,CAAC,KAAK,EAAE,MAAM,KAAK,IAAI,EAChC,UAAU,CAAC,EAAE,CAAC,QAAQ,EAAE,MAAM,EAAE,IAAI,EAAE,MAAM,KAAK,IAAI,GACpD,OAAO,CAAC,MAAM,CAAC,CAAA;IAElB,gBAAgB,CACd,MAAM,EAAE,MAAM,EACd,OAAO,EAAE,CAAC,SAAS,EAAE,MAAM,KAAK,IAAI,GACnC,OAAO,CAAC,MAAM,CAAC,CAAA;IAElB;;OAEG;IACH,IAAI,IAAI,IAAI,CAAA;IAEZ;;OAEG;IACH,MAAM,IAAI,IAAI,CAAA;IAEd;;;OAGG;IACH,sBAAsB,IAAI,eAAe,CAAA;IAEzC;;;OAGG;IACH,UAAU,IAAI,UAAU,EAAE,CAAA;IAE1B;;OAEG;IACH,YAAY,IAAI,IAAI,CAAA;IAEpB,0CAA0C;IAC1C,QAAQ,CAAC,QAAQ,EAAE,OAAO,CAAA;IAC1B,gDAAgD;IAChD,QAAQ,CAAC,YAAY,EAAE,OAAO,CAAA;IAC9B,2CAA2C;IAC3C,QAAQ,CAAC,OAAO,EAAE,MAAM,CAAA;IAExB,2BAA2B;IAC3B,KAAK,EAAE,OAAO,CAAA;IACd,gDAAgD;IAChD,YAAY,EAAE,MAAM,CAAA;IACpB,2CAA2C;IAC3C,SAAS,EAAE,MAAM,CAAA;IACjB,qEAAqE;IACrE,WAAW,EAAE,MAAM,CAAA;IACnB,sDAAsD;IACtD,cAAc,EAAE,OAAO,CAAA;CACxB;AAED;;;GAGG;AACH,MAAM,MAAM,iBAAiB,GAAG,QAAQ,GAAG,QAAQ,GAAG,SAAS,GAAG,OAAO,GAAG,QAAQ,CAAA"}

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@inferrlm/react-native-mlx",
   "description": "MLX Swift integration for React Native - InferrLM fork with enhanced features",
-  "version": "0.4.2-alpha.0",
+  "version": "0.4.2-alpha.10",
   "main": "./lib/module/index.js",
   "module": "./lib/module/index.js",
   "types": "./lib/typescript/src/index.d.ts",

package/src/llm.ts

@@ -226,4 +226,28 @@ export const LLM = {
   set systemPrompt(value: string) {
     getInstance().systemPrompt = value
   },
+
+  get maxTokens(): number {
+    return getInstance().maxTokens
+  },
+
+  set maxTokens(value: number) {
+    getInstance().maxTokens = value
+  },
+
+  get temperature(): number {
+    return getInstance().temperature
+  },
+
+  set temperature(value: number) {
+    getInstance().temperature = value
+  },
+
+  get enableThinking(): boolean {
+    return getInstance().enableThinking
+  },
+
+  set enableThinking(value: boolean) {
+    getInstance().enableThinking = value
+  },
 }

package/src/specs/LLM.nitro.ts

@@ -193,6 +193,12 @@ export interface LLM extends HybridObject<{ ios: 'swift' }> {
   debug: boolean
   /** System prompt used when loading the model */
   systemPrompt: string
+  /** Maximum number of tokens to generate */
+  maxTokens: number
+  /** Sampling temperature (0 = deterministic, higher = more random) */
+  temperature: number
+  /** Enable thinking mode for models that support it */
+  enableThinking: boolean
 }
 
 /**