@juspay/neurolink
Version:
Universal AI Development Platform with working MCP integration, multi-provider support, voice (TTS/STT/realtime), and professional CLI. 58+ external MCP servers discoverable, multimodal file processing, RAG pipelines. Build, test, and deploy AI applicatio
461 lines (460 loc) • 17.2 kB
JavaScript
/**
* Context Window Registry
*
* Accurate per-provider, per-model context window sizes (INPUT token limits).
* These are distinct from OUTPUT token limits in tokens.ts.
*
* Sources:
* - Anthropic: https://docs.anthropic.com/en/docs/about-claude/models
* - OpenAI: https://platform.openai.com/docs/models
* - Google: https://ai.google.dev/gemini-api/docs/models
* - Others: Provider documentation as of Feb 2026
*/
import { DynamicModelProvider } from "../core/dynamicModels.js";
import { logger } from "../utils/logger.js";
/** Default context window when provider/model is unknown */
export const DEFAULT_CONTEXT_WINDOW = 128_000;
/** Maximum output reserve when maxTokens not specified */
export const MAX_DEFAULT_OUTPUT_RESERVE = 64_000;
/** Default output reserve ratio (35% of context) */
export const DEFAULT_OUTPUT_RESERVE_RATIO = 0.35;
/**
* Per-provider, per-model context window sizes.
* The "_default" key is the fallback for unknown models within a provider.
*/
export const MODEL_CONTEXT_WINDOWS = {
deepseek: {
_default: 64_000,
"deepseek-chat": 64_000,
"deepseek-reasoner": 64_000,
},
"nvidia-nim": {
_default: 128_000,
"meta/llama-3.3-70b-instruct": 128_000,
"meta/llama-3.1-405b-instruct": 128_000,
"meta/llama-3.1-70b-instruct": 128_000,
"meta/llama-3.2-90b-vision-instruct": 128_000,
"meta/llama-3.2-11b-vision-instruct": 128_000,
"nvidia/llama-3.3-nemotron-super-49b-v1": 128_000,
"nvidia/llama-3.1-nemotron-nano-8b-v1": 128_000,
"nvidia/llama-3.1-nemotron-70b-instruct": 128_000,
"deepseek-ai/deepseek-r1": 128_000,
"deepseek-ai/deepseek-r1-distill-llama-70b": 128_000,
"mistralai/mixtral-8x22b-instruct-v0.1": 65_536,
"mistralai/mixtral-8x7b-instruct-v0.1": 32_768,
"microsoft/phi-4": 16_384,
"google/gemma-3-27b-it": 8_192,
},
"lm-studio": {
_default: 8_192,
},
llamacpp: {
_default: 8_192,
},
xai: {
_default: 131_072,
"grok-3": 131_072,
"grok-3-mini": 131_072,
"grok-2-latest": 131_072,
"grok-2-vision-latest": 32_768,
"grok-beta": 131_072,
},
groq: {
_default: 128_000,
"llama-3.3-70b-versatile": 131_072,
"llama-3.1-8b-instant": 128_000,
"llama-3.2-90b-vision-preview": 128_000,
"llama-3.2-11b-vision-preview": 128_000,
"llama-guard-3-8b": 8_192,
"gemma2-9b-it": 8_192,
"mixtral-8x7b-32768": 32_768,
},
cohere: {
_default: 128_000,
"command-r-plus": 128_000,
"command-r": 128_000,
"command-r7b-12-2024": 128_000,
},
"together-ai": {
_default: 128_000,
"meta-llama/Llama-3.3-70B-Instruct-Turbo": 128_000,
"meta-llama/Meta-Llama-3.1-405B-Instruct-Turbo": 128_000,
"meta-llama/Meta-Llama-3.1-70B-Instruct-Turbo": 128_000,
"meta-llama/Meta-Llama-3.1-8B-Instruct-Turbo": 128_000,
"mistralai/Mixtral-8x22B-Instruct-v0.1": 65_536,
"mistralai/Mixtral-8x7B-Instruct-v0.1": 32_768,
"Qwen/Qwen2.5-72B-Instruct-Turbo": 32_768,
"Qwen/Qwen2.5-Coder-32B-Instruct": 32_768,
"deepseek-ai/DeepSeek-R1": 64_000,
"deepseek-ai/DeepSeek-V3": 64_000,
"google/gemma-2-27b-it": 8_192,
"microsoft/WizardLM-2-8x22B": 65_536,
},
fireworks: {
_default: 128_000,
"accounts/fireworks/models/llama-v3p1-70b-instruct": 131_072,
"accounts/fireworks/models/llama-v3p1-405b-instruct": 128_000,
"accounts/fireworks/models/llama-v3p1-8b-instruct": 128_000,
"accounts/fireworks/models/llama-v3p3-70b-instruct": 128_000,
"accounts/fireworks/models/mixtral-8x22b-instruct": 65_536,
"accounts/fireworks/models/qwen2p5-72b-instruct": 32_768,
"accounts/fireworks/models/qwen2p5-coder-32b-instruct": 32_768,
"accounts/fireworks/models/deepseek-v3": 64_000,
},
perplexity: {
_default: 127_000,
sonar: 127_000,
"sonar-pro": 200_000,
"sonar-reasoning": 127_000,
"sonar-reasoning-pro": 127_000,
"sonar-deep-research": 200_000,
},
cloudflare: {
_default: 8_192,
"@cf/meta/llama-3.3-70b-instruct-fp8-fast": 24_000,
"@cf/meta/llama-3.1-70b-instruct": 24_000,
"@cf/meta/llama-3.1-8b-instruct-fast": 24_000,
"@cf/meta/llama-3.2-11b-vision-instruct": 24_000,
"@cf/mistral/mistral-7b-instruct-v0.2": 32_768,
"@cf/qwen/qwen1.5-14b-chat-awq": 7_500,
"@cf/google/gemma-2b-it-lora": 4_096,
},
replicate: {
// Per-model — Replicate hosts arbitrary models; sensible default.
_default: 32_768,
},
voyage: {
// Voyage embeddings: max input tokens vary 16K-32K per model
_default: 32_000,
"voyage-3.5": 32_000,
"voyage-3.5-lite": 32_000,
"voyage-3-large": 32_000,
"voyage-code-3": 32_000,
"voyage-finance-2": 32_000,
"voyage-law-2": 16_000,
"voyage-multilingual-2": 32_000,
},
jina: {
// Jina embeddings: 8K input tokens for v3; 8K for v2; 32K for ColBERT-v2
_default: 8_192,
"jina-embeddings-v3": 8_192,
"jina-embeddings-v2-base-en": 8_192,
"jina-embeddings-v2-small-en": 8_192,
"jina-colbert-v2": 32_000,
},
stability: {
// Image-gen — context is prompt length only; ~2000 char limit
_default: 2_000,
},
ideogram: {
_default: 2_000,
},
recraft: {
_default: 2_000,
},
anthropic: {
_default: 200_000,
// Claude 4.6 (Feb 2026) — 1M context window
"claude-opus-4-6": 1_000_000,
"claude-sonnet-4-6": 1_000_000,
// Claude 4.5
"claude-opus-4-5-20251101": 200_000,
"claude-sonnet-4-5-20250929": 200_000,
"claude-haiku-4-5-20251001": 200_000,
// Claude 4.x
"claude-opus-4-1-20250805": 200_000,
"claude-opus-4-20250514": 200_000,
"claude-sonnet-4-20250514": 200_000,
// Claude 3.x
"claude-3-7-sonnet-20250219": 200_000,
"claude-3-5-sonnet-20241022": 200_000,
"claude-3-5-haiku-20241022": 200_000,
"claude-3-opus-20240229": 200_000,
"claude-3-sonnet-20240229": 200_000,
"claude-3-haiku-20240307": 200_000,
},
openai: {
_default: 128_000,
// GPT-5.4 family — 1.05M context
"gpt-5.4": 1_050_000,
"gpt-5.4-mini": 400_000,
"gpt-5.4-nano": 400_000,
"gpt-5.4-pro": 1_050_000,
// GPT-5.x family — 400K context
"gpt-5.3-codex": 400_000,
"gpt-5.2": 400_000,
"gpt-5.2-pro": 400_000,
"gpt-5.2-codex": 400_000,
"gpt-5.2-chat-latest": 128_000,
"gpt-5.1": 400_000,
"gpt-5.1-codex": 400_000,
"gpt-5.1-codex-max": 400_000,
"gpt-5.1-codex-mini": 400_000,
"gpt-5.1-chat-latest": 128_000,
"gpt-5": 400_000,
"gpt-5-mini": 400_000,
"gpt-5-nano": 400_000,
"gpt-5-pro": 400_000,
"gpt-5-codex": 400_000,
"gpt-5-chat-latest": 128_000,
// GPT Open Source
"gpt-oss-120b": 128_000,
"gpt-oss-20b": 128_000,
// GPT-4.1 family — 1M context
"gpt-4.1": 1_047_576,
"gpt-4.1-mini": 1_047_576,
"gpt-4.1-nano": 1_047_576,
// GPT-4o
"gpt-4o": 128_000,
"gpt-4o-mini": 128_000,
// O-series reasoning — 200K context
o1: 200_000,
"o1-mini": 128_000,
"o1-pro": 200_000,
o3: 200_000,
"o3-mini": 200_000,
"o3-pro": 200_000,
"o4-mini": 200_000,
// Legacy
"gpt-4-turbo": 128_000,
"gpt-4": 8_192,
"gpt-3.5-turbo": 16_385,
},
"google-ai": {
_default: 1_048_576,
// Gemini 3.1 Series (all require -preview suffix)
"gemini-3.1-pro-preview": 1_048_576,
"gemini-3.1-flash-lite-preview": 1_048_576,
"gemini-3.1-flash-image-preview": 1_048_576,
"gemini-3.1-pro-preview-customtools": 1_048_576,
// Gemini 3 Series
"gemini-3-flash-preview": 1_048_576,
"gemini-3-pro-image-preview": 65_536,
/** @deprecated SHUT DOWN March 9, 2026. Migrate to gemini-3.1-pro-preview. */
"gemini-3-pro-preview": 1_048_576,
"gemini-2.5-pro": 1_048_576,
"gemini-2.5-flash": 1_048_576,
"gemini-2.5-flash-lite": 1_048_576,
"gemini-2.5-flash-image": 32_768,
"gemini-2.0-flash": 1_048_576,
"gemini-1.5-pro": 2_097_152,
"gemini-1.5-flash": 1_048_576,
},
vertex: {
_default: 1_048_576,
// Claude on Vertex
"claude-opus-4-6": 1_000_000,
"claude-sonnet-4-6": 1_000_000,
"claude-sonnet-4-5": 200_000,
"claude-opus-4-5": 200_000,
"claude-haiku-4-5": 200_000,
"claude-sonnet-4": 200_000,
"claude-sonnet-4-20250514": 200_000,
"claude-opus-4-20250514": 200_000,
"claude-opus-4": 200_000,
// Gemini 3.1 on Vertex (all require -preview suffix)
"gemini-3.1-pro-preview": 1_048_576,
"gemini-3.1-flash-lite-preview": 1_048_576,
"gemini-3.1-flash-image-preview": 1_048_576,
"gemini-3.1-pro-preview-customtools": 1_048_576,
// Gemini 3 on Vertex
"gemini-3-flash-preview": 1_048_576,
"gemini-3-pro-image-preview": 65_536,
/** @deprecated SHUT DOWN March 9, 2026. Migrate to gemini-3.1-pro-preview. */
"gemini-3-pro-preview": 1_048_576,
// Gemini 2.x on Vertex
"gemini-2.5-pro": 1_048_576,
"gemini-2.5-flash": 1_048_576,
"gemini-2.0-flash": 1_048_576,
"gemini-1.5-pro": 2_097_152,
"gemini-1.5-flash": 1_048_576,
},
bedrock: {
_default: 200_000,
// Claude 4.6
"anthropic.claude-opus-4-6-v1:0": 1_000_000,
"anthropic.claude-sonnet-4-6": 1_000_000,
// Claude 4.5
"anthropic.claude-opus-4-5-20251124-v1:0": 200_000,
"anthropic.claude-sonnet-4-5-20250929-v1:0": 200_000,
"anthropic.claude-haiku-4-5-20251001-v1:0": 200_000,
// Claude legacy
"anthropic.claude-3-5-sonnet-20241022-v1:0": 200_000,
"anthropic.claude-3-5-haiku-20241022-v1:0": 200_000,
"anthropic.claude-3-opus-20240229-v1:0": 200_000,
"anthropic.claude-3-sonnet-20240229-v1:0": 200_000,
"anthropic.claude-3-haiku-20240307-v1:0": 200_000,
// Amazon Nova
"amazon.nova-pro-v1:0": 300_000,
"amazon.nova-lite-v1:0": 300_000,
"amazon.nova-2-lite-v1:0": 1_000_000,
// Writer
"writer.palmyra-x5-v1:0": 1_000_000,
"writer.palmyra-x4-v1:0": 128_000,
// NVIDIA
"nvidia.nemotron-nano-3-30b": 256_000,
},
azure: {
_default: 128_000,
// GPT-5.4
"gpt-5.4": 1_050_000,
"gpt-5.4-mini": 400_000,
"gpt-5.4-nano": 400_000,
"gpt-5.4-pro": 1_050_000,
// GPT-5.x
"gpt-5.2": 400_000,
"gpt-5.2-pro": 400_000,
"gpt-5.2-codex": 400_000,
"gpt-5.1": 400_000,
"gpt-5": 400_000,
"gpt-5-mini": 400_000,
// GPT-4.1
"gpt-4.1": 1_047_576,
"gpt-4.1-mini": 1_047_576,
// GPT-4o
"gpt-4o": 128_000,
"gpt-4o-mini": 128_000,
// O-series
o3: 200_000,
"o3-mini": 200_000,
"o4-mini": 200_000,
// Legacy
"gpt-4-turbo": 128_000,
"gpt-4": 8_192,
},
mistral: {
_default: 128_000,
"mistral-large-latest": 256_000,
"mistral-large-2512": 256_000,
"mistral-medium-latest": 128_000,
"mistral-small-latest": 128_000,
"codestral-latest": 256_000,
"codestral-2508": 256_000,
"devstral-2512": 256_000,
"devstral-small-2512": 256_000,
"magistral-medium-latest": 128_000,
"mistral-small-2603": 256_000,
},
ollama: {
_default: 128_000,
},
litellm: {
_default: 128_000,
},
huggingface: {
_default: 32_000,
},
sagemaker: {
_default: 128_000,
// NVIDIA Nemotron 3 Nano (February 2026) — 1M context
"nvidia-nemotron-3-nano-30b": 1_000_000,
// Qwen3 VL — 32K context
"qwen3-vl-8b-instruct": 32_768,
},
};
/**
* Map of provider aliases to canonical MODEL_CONTEXT_WINDOWS keys.
*
* Callers reach `getContextWindowSize` via the unnormalized form on
* `options.provider` (e.g. CLI `--provider lmstudio`, alias `llama.cpp`),
* and `ProviderFactory.normalizeProviderName` runs only at instantiation —
* its output never reaches budget calculations. Without this normalization
* those alias forms miss the table and fall back to `DEFAULT_CONTEXT_WINDOW`,
* understating the budget for LM Studio / llama.cpp / NVIDIA NIM.
*
* The keys here are the result of stripping non-alpha characters, so
* `lm-studio` -> `lmstudio`, `nvidia-nim` -> `nvidianim`, `llama.cpp` -> `llamacpp`.
*/
const PROVIDER_ALIAS_MAP = {
googleaistudio: "google-ai-studio",
lmstudio: "lm-studio",
llamacpp: "llamacpp",
nvidianim: "nvidia-nim",
nim: "nvidia-nim",
nvidia: "nvidia-nim",
deepseek: "deepseek",
};
function normalizeProviderForLookup(provider) {
const stripped = provider.toLowerCase().replace(/[^a-z]/g, "");
// On alias miss, return the *stripped* key — not the raw input — so case /
// separator variants ("OpenAI", "open-ai", "Vertex AI") still find their
// table entry under the lowercase canonical key instead of falling through
// to DEFAULT_CONTEXT_WINDOW.
return PROVIDER_ALIAS_MAP[stripped] ?? stripped;
}
/**
* Resolve context window size for a provider/model combination.
*
* Priority:
* 0. Dynamic model registry (DynamicModelProvider) — resolves cross-provider
* models (e.g. Claude on Vertex) that the static table cannot handle
* 1. Exact model match under provider in static registry
* 2. Prefix match under provider in static registry
* 3. Provider's _default in static registry
* 4. Global DEFAULT_CONTEXT_WINDOW
*/
export function getContextWindowSize(provider, model) {
// Step 0: Check dynamic model registry first.
// This resolves cases where the runtime provider differs from the model's
// origin (e.g. Claude running via Vertex would hit Vertex's Gemini default
// in the static table). The dynamic registry knows the actual model metadata.
if (model) {
try {
const dynamicProvider = DynamicModelProvider.getInstance();
const modelConfig = dynamicProvider.resolveModel(provider, model);
if (modelConfig?.contextWindow) {
logger.debug(`[ContextWindow] Resolved via dynamic registry: provider=${provider}, model=${model}, contextWindow=${modelConfig.contextWindow}`);
return modelConfig.contextWindow;
}
}
catch {
// Dynamic registry not initialized yet — fall through to static lookup
}
}
// Static fallback chain — normalize aliases first so "lmstudio" / "llama.cpp" /
// "nvidianim" find their canonical entries instead of falling back to default.
const canonical = normalizeProviderForLookup(provider);
const providerWindows = MODEL_CONTEXT_WINDOWS[canonical] ?? MODEL_CONTEXT_WINDOWS[provider];
if (!providerWindows) {
return DEFAULT_CONTEXT_WINDOW;
}
if (model && providerWindows[model] !== undefined) {
return providerWindows[model];
}
// Try partial match (model name may be a prefix)
if (model) {
for (const [key, value] of Object.entries(providerWindows)) {
if (key !== "_default" && model.startsWith(key)) {
return value;
}
}
}
return providerWindows._default ?? DEFAULT_CONTEXT_WINDOW;
}
/**
* Calculate output token reserve for a given context window.
*
* Returns the *real* token count that will be reserved for output so callers
* (`getAvailableInputTokens`, `BudgetChecker`, conversation-memory pruning, file
* summarisation) compute input budget against the actual outgoing maxTokens.
*
* @param contextWindow - Total context window size
* @param maxTokens - Explicit maxTokens from user config (if set)
* @returns Number of tokens reserved for output (matches what's sent upstream)
*/
export function getOutputReserve(contextWindow, maxTokens) {
if (maxTokens !== undefined && maxTokens > 0) {
return maxTokens;
}
return Math.min(MAX_DEFAULT_OUTPUT_RESERVE, Math.ceil(contextWindow * DEFAULT_OUTPUT_RESERVE_RATIO));
}
/**
* Calculate available input tokens for a given provider/model.
*
* available = contextWindow - outputReserve
*/
export function getAvailableInputTokens(provider, model, maxTokens) {
const contextWindow = getContextWindowSize(provider, model);
const outputReserve = getOutputReserve(contextWindow, maxTokens);
return contextWindow - outputReserve;
}