@juspay/neurolink 12.0.4 → 12.1.0

package/dist/cli/commands/setup.js

@@ -20,7 +20,7 @@ import { handleGCPSetup } from "./setup-gcp.js";
 import { handleHuggingFaceSetup } from "./setup-huggingface.js";
 import { handleMistralSetup } from "./setup-mistral.js";
 import { PROVIDER_DESCRIPTORS_BY_NAME } from "../../factories/providerDescriptors.js";
-import { createCloudflareConfig, createCohereConfig, createDeepSeekConfig, createFireworksConfig, createGroqConfig, createIdeogramConfig, createJinaConfig, createNvidiaNimConfig, createOpenAICompatibleConfig, createPerplexityConfig, createRecraftConfig, createReplicateConfig, createStabilityConfig, createTogetherAIConfig, createVoyageConfig, createXaiConfig, satisfiesFallbacks, } from "../../utils/providerConfig.js";
+import { createCloudflareConfig, createCohereConfig, createDeepSeekConfig, createFireworksConfig, createCerebrasConfig, createGroqConfig, createIdeogramConfig, createJinaConfig, createNvidiaNimConfig, createOpenAICompatibleConfig, createPerplexityConfig, createRecraftConfig, createReplicateConfig, createStabilityConfig, createTogetherAIConfig, createVoyageConfig, createXaiConfig, satisfiesFallbacks, } from "../../utils/providerConfig.js";
 // Provider information database
 const PROVIDERS = [
     {
@@ -157,6 +157,7 @@ export const EXTRA_PROVIDER_CONFIGS = {
     "nvidia-nim": createNvidiaNimConfig(),
     xai: createXaiConfig(),
     groq: createGroqConfig(),
+    cerebras: createCerebrasConfig(),
     cohere: createCohereConfig(),
     replicate: createReplicateConfig(),
     "together-ai": createTogetherAIConfig(),

package/dist/constants/enums.d.ts

@@ -21,6 +21,7 @@ export declare enum AIProviderName {
     LLAMACPP = "llamacpp",
     XAI = "xai",
     GROQ = "groq",
+    CEREBRAS = "cerebras",
     COHERE = "cohere",
     TOGETHER_AI = "together-ai",
     FIREWORKS = "fireworks",
@@ -751,6 +752,24 @@ export declare enum GroqModels {
     /** Llama 3.2 11B Vision Preview — smaller multimodal */
     LLAMA_3_2_11B_VISION_PREVIEW = "llama-3.2-11b-vision-preview"
 }
+/**
+ * Cerebras inference models (wafer-scale, OpenAI-compatible API).
+ * @see https://inference-docs.cerebras.ai/introduction
+ *
+ * Note the vendor's inconsistent id scheme: `llama3.1-8b` has no dash
+ * after "llama", while `llama-3.3-70b` does — both are the DOCUMENTED
+ * ids, not typos.
+ */
+export declare enum CerebrasModels {
+    /** Llama 3.3 70B — production default */
+    LLAMA_3_3_70B = "llama-3.3-70b",
+    /** Llama 3.1 8B — low-latency tier (vendor id has no dash after "llama") */
+    LLAMA_3_1_8B = "llama3.1-8b",
+    /** Qwen 3 32B */
+    QWEN_3_32B = "qwen-3-32b",
+    /** OpenAI GPT-OSS 120B (open-weight) */
+    GPT_OSS_120B = "gpt-oss-120b"
+}
 /**
  * Cohere Command + Embed models.
  * @see https://docs.cohere.com/docs/models

package/dist/constants/enums.js

@@ -25,6 +25,7 @@ export var AIProviderName;
     AIProviderName["LLAMACPP"] = "llamacpp";
     AIProviderName["XAI"] = "xai";
     AIProviderName["GROQ"] = "groq";
+    AIProviderName["CEREBRAS"] = "cerebras";
     AIProviderName["COHERE"] = "cohere";
     AIProviderName["TOGETHER_AI"] = "together-ai";
     AIProviderName["FIREWORKS"] = "fireworks";
@@ -1004,6 +1005,25 @@ export var GroqModels;
     /** Llama 3.2 11B Vision Preview — smaller multimodal */
     GroqModels["LLAMA_3_2_11B_VISION_PREVIEW"] = "llama-3.2-11b-vision-preview";
 })(GroqModels || (GroqModels = {}));
+/**
+ * Cerebras inference models (wafer-scale, OpenAI-compatible API).
+ * @see https://inference-docs.cerebras.ai/introduction
+ *
+ * Note the vendor's inconsistent id scheme: `llama3.1-8b` has no dash
+ * after "llama", while `llama-3.3-70b` does — both are the DOCUMENTED
+ * ids, not typos.
+ */
+export var CerebrasModels;
+(function (CerebrasModels) {
+    /** Llama 3.3 70B — production default */
+    CerebrasModels["LLAMA_3_3_70B"] = "llama-3.3-70b";
+    /** Llama 3.1 8B — low-latency tier (vendor id has no dash after "llama") */
+    CerebrasModels["LLAMA_3_1_8B"] = "llama3.1-8b";
+    /** Qwen 3 32B */
+    CerebrasModels["QWEN_3_32B"] = "qwen-3-32b";
+    /** OpenAI GPT-OSS 120B (open-weight) */
+    CerebrasModels["GPT_OSS_120B"] = "gpt-oss-120b";
+})(CerebrasModels || (CerebrasModels = {}));
 /**
  * Cohere Command + Embed models.
  * @see https://docs.cohere.com/docs/models

package/dist/factories/providerDescriptors.js

@@ -1,5 +1,5 @@
 import { AIProviderName } from "../constants/enums.js";
-import { GoogleAIModels, OpenAIModels, AnthropicModels, VertexModels, MistralModels, OllamaModels, LiteLLMModels, HuggingFaceModels, DeepSeekModels, NvidiaNimModels, OpenRouterModels, XaiModels, GroqModels, CohereModels, TogetherAIModels, FireworksModels, PerplexityModels, CloudflareModels, VoyageModels, JinaModels, StabilityModels, IdeogramModels, RecraftModels, ReplicateModels, } from "../constants/enums.js";
+import { GoogleAIModels, OpenAIModels, AnthropicModels, VertexModels, MistralModels, OllamaModels, LiteLLMModels, HuggingFaceModels, DeepSeekModels, NvidiaNimModels, OpenRouterModels, XaiModels, GroqModels, CerebrasModels, CohereModels, TogetherAIModels, FireworksModels, PerplexityModels, CloudflareModels, VoyageModels, JinaModels, StabilityModels, IdeogramModels, RecraftModels, ReplicateModels, } from "../constants/enums.js";
 import { API_KEY_FORMATS } from "../utils/providerConfig.js";
 /**
  * Single source of truth for provider identity, credentials, defaults, and
@@ -368,6 +368,21 @@ export const PROVIDER_DESCRIPTORS = [
         healthCheck: "env-only",
         setupUrl: "https://console.x.ai/",
     },
+    {
+        name: AIProviderName.CEREBRAS,
+        aliases: [],
+        credentialsKey: "cerebras",
+        envVars: {
+            apiKey: "CEREBRAS_API_KEY",
+            baseURL: "CEREBRAS_BASE_URL",
+            model: "CEREBRAS_MODEL",
+        },
+        defaultModel: CerebrasModels.LLAMA_3_3_70B,
+        toolSupport: "native",
+        localRuntime: false,
+        healthCheck: "env-only",
+        setupUrl: "https://cloud.cerebras.ai",
+    },
     {
         name: AIProviderName.GROQ,
         aliases: [],

package/dist/models/manifestRegistry.js

@@ -18,6 +18,7 @@ import { lmStudioManifest } from "./manifests/lm-studio.js";
 import { llamacppManifest } from "./manifests/llamacpp.js";
 import { xaiManifest } from "./manifests/xai.js";
 import { groqManifest } from "./manifests/groq.js";
+import { cerebrasManifest } from "./manifests/cerebras.js";
 import { cohereManifest } from "./manifests/cohere.js";
 import { togetherAiManifest } from "./manifests/together-ai.js";
 import { fireworksManifest } from "./manifests/fireworks.js";
@@ -56,6 +57,7 @@ export const MANIFEST_REGISTRY = {
     llamacpp: llamacppManifest,
     xai: xaiManifest,
     groq: groqManifest,
+    cerebras: cerebrasManifest,
     cohere: cohereManifest,
     "together-ai": togetherAiManifest,
     fireworks: fireworksManifest,

package/dist/models/manifests/cerebras.d.ts

@@ -0,0 +1,9 @@
+import type { ProviderModelManifest } from "../../types/index.js";
+/**
+ * Minimal manifest: conservative floor values pending live verification —
+ * Cerebras serves large-context models, but the free tier caps effective
+ * context/output well below the architectural maximums, so these defaults
+ * stay deliberately modest. Named models can be added incrementally
+ * without touching any consumer — same pattern as groq.ts.
+ */
+export declare const cerebrasManifest: ProviderModelManifest;

package/dist/models/manifests/cerebras.js

@@ -0,0 +1,19 @@
+/**
+ * Minimal manifest: conservative floor values pending live verification —
+ * Cerebras serves large-context models, but the free tier caps effective
+ * context/output well below the architectural maximums, so these defaults
+ * stay deliberately modest. Named models can be added incrementally
+ * without touching any consumer — same pattern as groq.ts.
+ */
+export const cerebrasManifest = {
+    defaultContextWindow: 65536,
+    models: {
+        _default: {
+            aliases: [],
+            contextWindow: 65536,
+            maxOutputTokens: 8192,
+            vision: false,
+            functionCalling: true,
+        },
+    },
+};

package/dist/neurolink.js

@@ -6070,32 +6070,118 @@ Current user's request: ${currentInput}`;
                 // oversized case. When the budget check shows the request is
                 // over budget but there's nothing to compact (no memory + no
                 // inline messages — e.g. a huge prompt or huge tool definitions
-                // alone), throw before dispatch instead of wasting a roundtrip.
+                // alone), recover by WINDOWING the prompt when the prompt is what
+                // blew the budget: keep its head and tail around an elision marker
+                // and dispatch. Agentic callers (Yama's session loop) carry their
+                // whole tool transcript in the prompt; the previous
+                // unconditional throw dead-ended every such turn — observed live
+                // as an unbounded retry loop (93K→299K tokens, no verdict, ever).
+                // The throw remains for the truly unrecoverable case: system
+                // prompt + tool definitions alone exceed the budget.
+                let promptWindowRecovered = false;
                 if (!budgetCheck.withinBudget && !dpgHasCompactableMessages) {
-                    try {
-                        this.emitter.emit("compaction.insufficient", {
-                            stagesAttempted: ["pre-dispatch hard cap"],
-                            finalTokens: budgetCheck.estimatedInputTokens,
-                            budget: budgetCheck.availableInputTokens,
-                            provider: providerName,
-                            model: options.model,
-                            phase: "pre-dispatch-no-recovery",
-                            timestamp: Date.now(),
-                        });
+                    const fixedOverhead = (budgetCheck.breakdown?.systemPrompt ?? 0) +
+                        (budgetCheck.breakdown?.toolDefinitions ?? 0) +
+                        (budgetCheck.breakdown?.fileAttachments ?? 0);
+                    // 3% margin against estimator drift; 1024-token floor — below
+                    // that, a windowed prompt carries too little to answer from.
+                    const promptBudget = Math.floor((budgetCheck.availableInputTokens - fixedOverhead) * 0.97);
+                    const promptText = typeof options.prompt === "string" ? options.prompt : undefined;
+                    if (promptText &&
+                        promptBudget >= 1024 &&
+                        (budgetCheck.breakdown?.currentPrompt ?? 0) > promptBudget) {
+                        const marker = "\n\n[... middle of this prompt elided by NeuroLink to fit the model's context window ...]\n\n";
+                        // Proportional char budget from the observed chars-per-token of
+                        // THIS text, re-checked and shrunk until the estimator agrees.
+                        let charBudget = Math.floor(promptText.length *
+                            (promptBudget /
+                                Math.max(budgetCheck.breakdown?.currentPrompt ?? 1, 1)));
+                        let windowed = promptText;
+                        for (let attempt = 0; attempt < 4; attempt++) {
+                            const headChars = Math.floor(charBudget * 0.6);
+                            const tailChars = Math.max(charBudget - headChars - marker.length, 0);
+                            windowed =
+                                promptText.slice(0, headChars) +
+                                    marker +
+                                    (tailChars > 0 ? promptText.slice(-tailChars) : "");
+                            const recheck = checkContextBudget({
+                                provider: providerName,
+                                model: options.model,
+                                maxTokens: options.maxTokens,
+                                systemPrompt: options.systemPrompt,
+                                conversationMessages: [],
+                                currentPrompt: windowed,
+                                toolDefinitions: options.tools
+                                    ? Object.values(options.tools)
+                                    : undefined,
+                            });
+                            if (recheck.withinBudget) {
+                                break;
+                            }
+                            charBudget = Math.floor(charBudget * 0.8);
+                            if (attempt === 3) {
+                                windowed = promptText; // give up — fall through to the throw
+                            }
+                        }
+                        if (windowed !== promptText) {
+                            logger.warn("[NeuroLink] Prompt exceeded the model's context budget with nothing to compact — " +
+                                "windowed the prompt (head+tail kept, middle elided) to fit.", {
+                                provider: providerName,
+                                model: options.model,
+                                estimatedTokens: budgetCheck.estimatedInputTokens,
+                                budget: budgetCheck.availableInputTokens,
+                                originalPromptChars: promptText.length,
+                                windowedPromptChars: windowed.length,
+                            });
+                            try {
+                                this.emitter.emit("compaction.applied", {
+                                    stagesAttempted: ["pre-dispatch prompt window"],
+                                    finalTokens: budgetCheck.availableInputTokens,
+                                    budget: budgetCheck.availableInputTokens,
+                                    provider: providerName,
+                                    model: options.model,
+                                    phase: "pre-dispatch-prompt-window",
+                                    timestamp: Date.now(),
+                                });
+                            }
+                            catch {
+                                /* listener errors are non-fatal */
+                            }
+                            options.prompt = windowed;
+                            const inputHolder = options
+                                .input;
+                            if (inputHolder && typeof inputHolder.text === "string") {
+                                inputHolder.text = windowed;
+                            }
+                            promptWindowRecovered = true;
+                        }
                     }
-                    catch {
-                        /* listener errors are non-fatal */
+                    if (!promptWindowRecovered) {
+                        try {
+                            this.emitter.emit("compaction.insufficient", {
+                                stagesAttempted: ["pre-dispatch hard cap"],
+                                finalTokens: budgetCheck.estimatedInputTokens,
+                                budget: budgetCheck.availableInputTokens,
+                                provider: providerName,
+                                model: options.model,
+                                phase: "pre-dispatch-no-recovery",
+                                timestamp: Date.now(),
+                            });
+                        }
+                        catch {
+                            /* listener errors are non-fatal */
+                        }
+                        throw new ContextBudgetExceededError(`Context exceeds model budget and no compaction is possible ` +
+                            `(no conversationMemory, no inline conversationMessages — only ` +
+                            `prompt + tools). Estimated: ${budgetCheck.estimatedInputTokens} ` +
+                            `tokens, budget: ${budgetCheck.availableInputTokens} tokens. ` +
+                            `Reduce prompt or tool-definition size, or trim the request.`, {
+                            estimatedTokens: budgetCheck.estimatedInputTokens,
+                            availableTokens: budgetCheck.availableInputTokens,
+                            stagesUsed: [],
+                            breakdown: budgetCheck.breakdown,
+                        });
                     }
-                    throw new ContextBudgetExceededError(`Context exceeds model budget and no compaction is possible ` +
-                        `(no conversationMemory, no inline conversationMessages — only ` +
-                        `prompt + tools). Estimated: ${budgetCheck.estimatedInputTokens} ` +
-                        `tokens, budget: ${budgetCheck.availableInputTokens} tokens. ` +
-                        `Reduce prompt or tool-definition size, or trim the request.`, {
-                        estimatedTokens: budgetCheck.estimatedInputTokens,
-                        availableTokens: budgetCheck.availableInputTokens,
-                        stagesUsed: [],
-                        breakdown: budgetCheck.breakdown,
-                    });
                 }
                 if (budgetCheck.shouldCompact &&
                     (this.conversationMemory || dpgHasInlineMessages) &&

package/dist/providers/openaiCompatCatalog.d.ts

@@ -1,6 +1,6 @@
 import type { OpenAICompatCatalogEntry } from "../types/index.js";
 /**
- * Config-driven catalog of the 7 zero-quirk OpenAI-compatible providers.
+ * Config-driven catalog of the 8 zero-quirk OpenAI-compatible providers.
  * Each entry fully replaces what used to be a hand-written
  * OpenAIChatCompletionsProvider subclass — see ConfiguredOpenAICompatProvider
  * for the class that reads these entries, and providerRegistry.ts for the

package/dist/providers/openaiCompatCatalog.js

@@ -1,13 +1,13 @@
 import { AIProviderName } from "../constants/enums.js";
-import { CloudflareModels, FireworksModels, GroqModels, MistralModels, PerplexityModels, TogetherAIModels, XaiModels, } from "../constants/enums.js";
+import { CerebrasModels, CloudflareModels, FireworksModels, GroqModels, MistralModels, PerplexityModels, TogetherAIModels, XaiModels, } from "../constants/enums.js";
 import { AuthenticationError, InvalidModelError, ProviderError, } from "../types/index.js";
 import { DEFAULT_ERROR_RULES } from "../utils/errorClassifier.js";
-import { createCloudflareConfig, createFireworksConfig, createGroqConfig, createMistralConfig, createPerplexityConfig, createTogetherAIConfig, createXaiConfig, } from "../utils/providerConfig.js";
+import { createCerebrasConfig, createCloudflareConfig, createFireworksConfig, createGroqConfig, createMistralConfig, createPerplexityConfig, createTogetherAIConfig, createXaiConfig, } from "../utils/providerConfig.js";
 function buildCloudflareBaseURL(accountId) {
     return `https://api.cloudflare.com/client/v4/accounts/${accountId}/ai/v1`;
 }
 /**
- * Config-driven catalog of the 7 zero-quirk OpenAI-compatible providers.
+ * Config-driven catalog of the 8 zero-quirk OpenAI-compatible providers.
  * Each entry fully replaces what used to be a hand-written
  * OpenAIChatCompletionsProvider subclass — see ConfiguredOpenAICompatProvider
  * for the class that reads these entries, and providerRegistry.ts for the
@@ -29,6 +29,37 @@ function buildCloudflareBaseURL(accountId) {
  * Task 14's docs task for the deciding criteria).
  */
 export const OPENAI_COMPAT_CATALOG = [
+    {
+        providerName: AIProviderName.CEREBRAS,
+        aliases: ["cerebras"],
+        apiKeyEnvVar: "CEREBRAS_API_KEY",
+        baseURLEnvVar: "CEREBRAS_BASE_URL",
+        defaultBaseURL: "https://api.cerebras.ai/v1",
+        configOptions: createCerebrasConfig(),
+        modelEnvVar: "CEREBRAS_MODEL",
+        defaultModel: CerebrasModels.LLAMA_3_3_70B,
+        registryDefaultModel: CerebrasModels.LLAMA_3_3_70B,
+        registryDefaultModelChecksEnvVar: true,
+        fallbackModelName: CerebrasModels.LLAMA_3_1_8B,
+        fallbackModels: [
+            CerebrasModels.LLAMA_3_3_70B,
+            CerebrasModels.LLAMA_3_1_8B,
+            CerebrasModels.QWEN_3_32B,
+            CerebrasModels.GPT_OSS_120B,
+        ],
+        errorRules: [
+            {
+                // Probed live 2026-08-26: a bad key gets HTTP 401 with body
+                // {"message":"Wrong API Key","type":"invalid_request_error",
+                //  "param":"api_key","code":"wrong_api_key"}.
+                match: (ctx) => ctx.statusCode === 401 ||
+                    /wrong_api_key|Wrong API Key|invalid_api_key/i.test(ctx.message),
+                errorClass: AuthenticationError,
+                message: "Invalid Cerebras API key. Check CEREBRAS_API_KEY. Get one at https://cloud.cerebras.ai",
+            },
+            ...DEFAULT_ERROR_RULES,
+        ],
+    },
     {
         providerName: AIProviderName.GROQ,
         aliases: ["groq"],

package/dist/types/providers.d.ts

@@ -178,6 +178,10 @@ export type NeurolinkCredentials = {
         apiKey?: string;
         baseURL?: string;
     };
+    cerebras?: {
+        apiKey?: string;
+        baseURL?: string;
+    };
     cohere?: {
         apiKey?: string;
         baseURL?: string;

package/dist/utils/modelChoices.js

@@ -2,7 +2,7 @@
  * Centralized model choices for CLI commands
  * Derives choices from model enums to ensure consistency
  */
-import { AIProviderName, OpenAIModels, AnthropicModels, GoogleAIModels, BedrockModels, VertexModels, MistralModels, OllamaModels, AzureOpenAIModels, LiteLLMModels, HuggingFaceModels, SageMakerModels, OpenRouterModels, DeepSeekModels, NvidiaNimModels, XaiModels, GroqModels, CohereModels, TogetherAIModels, FireworksModels, PerplexityModels, CloudflareModels, VoyageModels, JinaModels, StabilityModels, IdeogramModels, RecraftModels, ReplicateModels, } from "../constants/enums.js";
+import { AIProviderName, OpenAIModels, AnthropicModels, GoogleAIModels, BedrockModels, VertexModels, MistralModels, OllamaModels, AzureOpenAIModels, LiteLLMModels, HuggingFaceModels, SageMakerModels, OpenRouterModels, DeepSeekModels, NvidiaNimModels, XaiModels, GroqModels, CerebrasModels, CohereModels, TogetherAIModels, FireworksModels, PerplexityModels, CloudflareModels, VoyageModels, JinaModels, StabilityModels, IdeogramModels, RecraftModels, ReplicateModels, } from "../constants/enums.js";
 /**
  * Top models per provider with descriptions for CLI prompts
  * These are curated lists of the most commonly used/recommended models
@@ -302,6 +302,21 @@ const TOP_MODELS_CONFIG = {
             description: "Mistral 8x7B MoE, 32K context",
         },
     ],
+    [AIProviderName.CEREBRAS]: [
+        {
+            model: CerebrasModels.LLAMA_3_3_70B,
+            description: "Recommended - Production default; wafer-scale speed",
+        },
+        {
+            model: CerebrasModels.LLAMA_3_1_8B,
+            description: "Lowest latency tier",
+        },
+        { model: CerebrasModels.QWEN_3_32B, description: "Qwen 3 32B" },
+        {
+            model: CerebrasModels.GPT_OSS_120B,
+            description: "OpenAI GPT-OSS 120B (open-weight)",
+        },
+    ],
     [AIProviderName.COHERE]: [
         {
             model: CohereModels.COMMAND_R_PLUS,
@@ -539,6 +554,7 @@ const MODEL_ENUMS = {
     [AIProviderName.LLAMACPP]: null,
     [AIProviderName.XAI]: XaiModels,
     [AIProviderName.GROQ]: GroqModels,
+    [AIProviderName.CEREBRAS]: CerebrasModels,
     [AIProviderName.COHERE]: CohereModels,
     [AIProviderName.TOGETHER_AI]: TogetherAIModels,
     [AIProviderName.FIREWORKS]: FireworksModels,

package/dist/utils/providerConfig.d.ts

@@ -135,6 +135,10 @@ export declare function createNvidiaNimConfig(): ProviderConfigOptions;
  * Creates xAI Grok provider configuration.
  */
 export declare function createXaiConfig(): ProviderConfigOptions;
+/**
+ * Creates Cerebras provider configuration.
+ */
+export declare function createCerebrasConfig(): ProviderConfigOptions;
 /**
  * Creates Groq provider configuration.
  */

package/dist/utils/providerConfig.js

@@ -451,6 +451,23 @@ export function createXaiConfig() {
         ],
     };
 }
+/**
+ * Creates Cerebras provider configuration.
+ */
+export function createCerebrasConfig() {
+    return {
+        providerName: "Cerebras",
+        envVarName: "CEREBRAS_API_KEY",
+        setupUrl: "https://cloud.cerebras.ai",
+        description: "API key",
+        instructions: [
+            "1. Visit: https://cloud.cerebras.ai",
+            "2. Sign in or create a free Cerebras account",
+            "3. Create an API key under API Keys",
+            "4. Set CEREBRAS_API_KEY in your .env file",
+        ],
+    };
+}
 /**
  * Creates Groq provider configuration.
  */

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@juspay/neurolink",
-  "version": "12.0.4",
+  "version": "12.1.0",
   "packageManager": "pnpm@10.15.1",
   "description": "TypeScript AI SDK with 24+ LLM providers behind one consistent API. MCP-native (connect any MCP server), voice TTS/STT/realtime, RAG, agents, memory, context compaction. OpenAI · Anthropic · Gemini · Bedrock · Azure · Ollama · DeepSeek · NVIDIA NIM and more.",
   "author": {