Press n or j to go to the next uncovered block, b, p or k for the previous block.
| 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 | 7x 7x 7x 7x 7x 7x 7x 7x 7x 130x 6x 124x 8x 116x 110x 6x 151x 151x 151x 129x 142x 154x 154x 12x 117x 127x 127x 127x 127x 127x 4x 127x 127x 127x 127x 127x 127x 127x 127x 127x 127x 127x 127x 127x 127x 127x 127x 127x 11x 11x 127x 127x 127x 127x 127x 127x 4x 4x 4x 3x 3x 3x 3x 3x 3x 127x 127x 127x 4x 4x 3x 3x 3x 2x 2x 2x 2x 2x 2x 124x 124x 124x 124x 124x 124x 124x 124x 124x 110x 124x 124x 1x 124x 119x 124x 124x 124x 124x 124x 1x 124x 124x 124x 124x 124x 11x 124x 11x 124x 124x 124x 124x 124x 124x | import * as vscode from "vscode";
import { ConfigManager } from "../shared/config";
import { calculateSafetyMargin, DEFAULT_MAX_OUTPUT_TOKENS } from "../shared/constants";
import {
convertMessages,
convertTools,
estimateMessagesTokens,
estimateNimMessagesTokens,
estimateToolsTokens,
truncateMessagesForContext,
LegacyPart,
} from "../messages/converter";
import { splitMessagesForSummarization, summarizeOldMessages } from "../models/summarizer";
import { getModelAdapter } from "../models/adapters";
import { createStructuredError } from "../api/errors";
import { debugLog } from "../shared/logging";
import { NimChatRequest, NimChatMessage, NimTool } from "../types";
export interface PreparedRequest {
requestBody: NimChatRequest;
reasoningIsolationExpected: boolean;
inputTokenCount: number;
requestedMaxTokens: number;
safetyMargin: number;
temperatureVal: number;
toolsEnabled: boolean;
extraSystemMessages: string[];
tools?: NimTool[];
}
/**
* When the prepared payload exceeds this fraction of the effective input budget,
* proactively compact older turns before the hard limit is reached. This
* avoids the server rejecting the request outright and gives the user a
* smoother experience in long conversations.
*/
const PREFLIGHT_COMPACT_THRESHOLD = 0.85;
export class NimRequestBuilder {
public static calculateMaxToolResultChars(contextWindow: number): number {
if (contextWindow >= 500000) {
return 50000;
}
if (contextWindow >= 200000) {
return 30000;
}
if (contextWindow >= 100000) {
return 20000;
}
return 10000;
}
public static calculateRequestedMaxTokens(options: {
requestedMaxTokens: number;
modelMaxOutputTokens: number;
contextWindow: number;
inputTokenCount: number;
}): number {
const safetyMargin = calculateSafetyMargin(options.contextWindow);
const availableCompletionTokens = Math.max(
1,
options.contextWindow - options.inputTokenCount - safetyMargin,
);
return Math.min(
options.requestedMaxTokens,
options.modelMaxOutputTokens,
availableCompletionTokens,
);
}
public static hasImageInput(messages: readonly vscode.LanguageModelChatMessage[]): boolean {
for (const msg of messages) {
for (const part of msg.content) {
const p = part as { mimeType?: unknown; data?: unknown };
if (typeof p.mimeType === "string" && p.mimeType.startsWith("image/")) {
return true;
}
}
}
return false;
}
public static async prepareRequest(options: {
model: vscode.LanguageModelChatInformation;
messages: readonly vscode.LanguageModelChatMessage[];
options: vscode.ProvideLanguageModelChatResponseOptions;
contextWindow: number;
supportsTools: boolean;
supportsVision: boolean;
apiKey: string;
userAgent: string;
signal?: AbortSignal;
}): Promise<PreparedRequest> {
const {
model,
messages,
options: responseOptions,
contextWindow,
supportsTools,
supportsVision,
apiKey,
userAgent,
signal,
} = options;
const rawInputTokenCount = estimateMessagesTokens(
messages as readonly { content: (vscode.LanguageModelInputPart | LegacyPart)[] }[],
);
const maxInputTokens = model.maxInputTokens;
const effectiveMaxInputTokens = Math.max(
1,
maxInputTokens - calculateSafetyMargin(contextWindow),
);
if (rawInputTokenCount > effectiveMaxInputTokens) {
debugLog(
"contextCompression",
`Input tokens ${rawInputTokenCount} exceed max ${effectiveMaxInputTokens}. Will attempt context compression.`,
);
}
const generationConfig = ConfigManager.getGenerationConfig();
const reasoningConfig = ConfigManager.getReasoningConfig();
const maxTokensVal = model.maxOutputTokens;
const modelMaxLimit =
typeof maxTokensVal === "number" && maxTokensVal > 0
? maxTokensVal
: DEFAULT_MAX_OUTPUT_TOKENS;
const requestedMaxTokensLimit =
typeof generationConfig.maxOutputTokens === "number" && generationConfig.maxOutputTokens > 0
? Math.min(modelMaxLimit, generationConfig.maxOutputTokens)
: modelMaxLimit;
const maxToolResultChars = this.calculateMaxToolResultChars(contextWindow);
const toolConfig = supportsTools ? convertTools(responseOptions) : {};
const toolsEnabled = Boolean(toolConfig.tools?.length);
const adapter = getModelAdapter(model.id);
const requestProfile = adapter.getProfile({
toolsEnabled,
});
const userTemperature = (responseOptions.modelOptions as Record<string, unknown>)?.temperature;
const profileTemperature =
toolsEnabled && requestProfile.toolTemperature !== undefined
? requestProfile.toolTemperature
: requestProfile.defaultTemperature;
const configTemperature = generationConfig.temperature;
const temperatureVal =
typeof userTemperature === "number"
? userTemperature
: typeof configTemperature === "number"
? configTemperature
: profileTemperature;
let apiMessages = convertMessages(messages, {
maxToolResultChars,
supportsVision,
});
apiMessages = adapter.applyMessagesWorkaround
? adapter.applyMessagesWorkaround(apiMessages)
: apiMessages;
if (requestProfile.extraSystemMessages.length > 0) {
apiMessages = [
...requestProfile.extraSystemMessages.map(
(content): NimChatMessage => ({ role: "system", content }),
),
...apiMessages,
];
}
const toolDefinitionTokens = toolConfig.tools ? estimateToolsTokens(toolConfig.tools) : 0;
let apiTokenCount = estimateNimMessagesTokens(apiMessages);
let payloadInputTokenCount = apiTokenCount + toolDefinitionTokens;
const messageTokenBudget = Math.max(1, effectiveMaxInputTokens - toolDefinitionTokens);
const recentTokenBudget = Math.floor(messageTokenBudget * 0.4);
/**
* Summarise older turns and replace them with a compact summary message.
* Returns the new messages array and updated payloadInputTokenCount.
*/
const compactMessages = async (
currentMessages: NimChatMessage[],
label: string,
): Promise<{ messages: NimChatMessage[]; tokenCount: number }> => {
const { oldMessages, recentMessages } = splitMessagesForSummarization(
currentMessages,
recentTokenBudget,
);
Iif (oldMessages.length === 0) {
return { messages: currentMessages, tokenCount: payloadInputTokenCount };
}
const summaryMessage = await summarizeOldMessages(
oldMessages,
apiKey,
userAgent,
signal,
ConfigManager.getContextConfig().summarizationModel,
);
const recentSystemMessages = recentMessages.filter((m) => m.role === "system");
const recentConversationMessages = recentMessages.filter((m) => m.role !== "system");
const compacted = [...recentSystemMessages, summaryMessage, ...recentConversationMessages];
const tokenCount = estimateNimMessagesTokens(compacted) + toolDefinitionTokens;
debugLog(
"contextCompression",
`${label}: ${tokenCount} tokens (was ${payloadInputTokenCount}).`,
);
return { messages: compacted, tokenCount };
};
// --- Preflight compaction: proactive compression before hard limit ---
// When the payload exceeds the threshold (default 85%), compact older
// turns preemptively to avoid a server-side 400 rejection.
const compactThreshold = Math.floor(effectiveMaxInputTokens * PREFLIGHT_COMPACT_THRESHOLD);
Iif (
payloadInputTokenCount > compactThreshold &&
payloadInputTokenCount <= effectiveMaxInputTokens
) {
debugLog(
"contextCompression",
`Preflight: ${payloadInputTokenCount} tokens >= ${compactThreshold} threshold (${(PREFLIGHT_COMPACT_THRESHOLD * 100).toFixed(0)}%). Compacting proactively...`,
);
const result = await compactMessages(apiMessages, "Preflight compaction");
apiMessages = result.messages;
payloadInputTokenCount = result.tokenCount;
}
// --- Hard limit: reactive compression when over budget ---
if (payloadInputTokenCount > effectiveMaxInputTokens) {
debugLog(
"contextCompression",
`Prepared payload ${payloadInputTokenCount} tokens > ${effectiveMaxInputTokens} max. Compressing...`,
);
const result = await compactMessages(apiMessages, "Hard-limit compaction");
apiMessages = result.messages;
payloadInputTokenCount = result.tokenCount;
if (payloadInputTokenCount > effectiveMaxInputTokens) {
apiMessages = truncateMessagesForContext(apiMessages, messageTokenBudget);
const finalMessageTokenCount = estimateNimMessagesTokens(apiMessages);
payloadInputTokenCount = finalMessageTokenCount + toolDefinitionTokens;
debugLog(
"contextCompression",
`After truncation fallback: ${payloadInputTokenCount} tokens.`,
);
if (payloadInputTokenCount > effectiveMaxInputTokens) {
throw createStructuredError(
"token_limit",
`Even after compression and truncation: ${payloadInputTokenCount} tokens, max: ${effectiveMaxInputTokens}`,
);
}
}
}
apiTokenCount = estimateNimMessagesTokens(apiMessages);
payloadInputTokenCount = apiTokenCount + toolDefinitionTokens;
Iif (payloadInputTokenCount > effectiveMaxInputTokens) {
throw createStructuredError(
"token_limit",
`Prepared payload exceeds context after compression: ${payloadInputTokenCount} tokens, max: ${effectiveMaxInputTokens}`,
);
}
const requestedMaxTokens = this.calculateRequestedMaxTokens({
requestedMaxTokens: requestedMaxTokensLimit,
modelMaxOutputTokens: model.maxOutputTokens,
contextWindow,
inputTokenCount: payloadInputTokenCount,
});
const requestBody: NimChatRequest = {
model: model.id,
messages: apiMessages,
stream: true,
max_tokens: requestedMaxTokens,
temperature: temperatureVal,
stream_options: { include_usage: true },
};
const configuredReasoningMode = (
responseOptions as { modelConfiguration?: { reasoningMode?: string } }
).modelConfiguration?.reasoningMode;
const modes = adapter.supportedReasoningModes;
let reasoningMode = configuredReasoningMode;
if (reasoningMode === undefined && modes && modes.length > 0) {
reasoningMode = reasoningConfig.mode;
}
reasoningMode ??= "none";
if (modes && modes.length > 0 && !modes.includes(reasoningMode)) {
reasoningMode = modes.includes("none") ? "none" : modes[0];
}
if (adapter.applyReasoningMode) {
adapter.applyReasoningMode(requestBody, reasoningMode);
}
const reasoningContentExpected =
Boolean(adapter.applyReasoningMode) && reasoningMode !== "none";
const reasoningIsolationExpected =
(reasoningContentExpected && adapter.isolateUntaggedReasoning !== false) ||
Boolean(adapter.alwaysReasons);
const modelOpts = responseOptions.modelOptions as Record<string, unknown>;
Iif (typeof modelOpts?.top_p === "number") {
requestBody.top_p = Math.min(1, Math.max(0, modelOpts.top_p));
} else if (typeof generationConfig.topP === "number") {
requestBody.top_p = Math.min(1, Math.max(0, generationConfig.topP));
}
Iif (typeof modelOpts?.frequency_penalty === "number") {
requestBody.frequency_penalty = Math.min(2, Math.max(-2, modelOpts.frequency_penalty));
}
Iif (typeof modelOpts?.presence_penalty === "number") {
requestBody.presence_penalty = Math.min(2, Math.max(-2, modelOpts.presence_penalty));
}
const stopVal = modelOpts?.stop;
Iif (typeof stopVal === "string" || (Array.isArray(stopVal) && stopVal.length > 0)) {
requestBody.stop = stopVal as string | string[];
}
if (toolConfig.tools) {
requestBody.tools = toolConfig.tools;
}
if (toolConfig.tool_choice) {
requestBody.tool_choice = toolConfig.tool_choice;
}
debugLog("Outgoing request messages", requestBody.messages);
const safetyMargin = calculateSafetyMargin(contextWindow);
const remainingBudget = Math.max(
0,
contextWindow - payloadInputTokenCount - requestedMaxTokens - safetyMargin,
);
const utilizationPercent =
contextWindow > 0 ? ((payloadInputTokenCount / contextWindow) * 100).toFixed(1) : "0";
debugLog("budget", {
contextWindow,
safetyMargin,
estimatedInputTokens: payloadInputTokenCount,
reservedOutputTokens: requestedMaxTokens,
remainingBudget,
utilizationPercent: `${utilizationPercent}%`,
toolDefinitionTokens,
});
return {
requestBody,
reasoningIsolationExpected,
inputTokenCount: payloadInputTokenCount,
requestedMaxTokens,
safetyMargin,
temperatureVal,
toolsEnabled,
extraSystemMessages: requestProfile.extraSystemMessages,
tools: toolConfig.tools,
};
}
}
|