@juspay/neurolink 12.38.0 → 12.39.0

package/dist/providers/catalog/neuralwatt.json

@@ -0,0 +1,179 @@
+{
+    "$schema": "./provider-catalog.schema.json",
+    "id": "neuralwatt",
+    "displayName": "Neuralwatt",
+    "aliases": [],
+    "tier": 2,
+    "wire": {
+        "baseURL": "https://api.neuralwatt.com/v1"
+    },
+    "models": {
+        "default": "glm-5.3",
+        "fallbacks": [
+            "glm-5.3-flash",
+            "qwen-3.8-27b",
+            "qwen3.6-35b",
+            "kimi-k2.7-code",
+            "kimi-k3",
+            "deepseek-v4.1-flash",
+            "gemma-4-31b"
+        ],
+        "defaultContextWindow": 1048560,
+        "defaultMaxOutputTokens": 4096,
+        "visionModel": "qwen-3.8-27b",
+        "catalog": {
+            "glm-5.3": {
+                "contextWindow": 1048560,
+                "pricingPerMTok": {
+                    "input": 1.45,
+                    "output": 4.5,
+                    "cachedInput": 0.145
+                },
+                "vision": false,
+                "status": "production",
+                "description": "ZhipuAI GLM 5.3 (fp8-native, zai-org/GLM-5.3) — the model used in every quickstart code sample on https://docs.neuralwatt.com/quickstart (curl, Python and Node.js all default to \"glm-5.3\"), so it is treated as Neuralwatt's own recommended default here. Roster metadata (GET /v1/models, 2026-09-28) marks reasoning mandatory (default effort \"max\") and declares tools:true and json_mode:true, but neither claim was independently probed — this onboarding is fully credential-free, no POST request was made. json_mode is not backed by a documented `response_format` parameter anywhere in the chat-completions API reference (https://docs.neuralwatt.com/api-reference/inference/create-a-chat-completion, retrieved 2026-09-28), so capabilities.structuredOutput is left false rather than trusting the flag alone. limits.max_output_tokens is null for this model in the live roster — omitted here rather than invented; models.defaultMaxOutputTokens (4096) is a generic placeholder, not a Neuralwatt-stated ceiling for this model. Pricing per the same roster response: $1.45 / $4.50 per MTok input/output, $0.145 cached input."
+            },
+            "glm-5.3-flash": {
+                "contextWindow": 1048560,
+                "pricingPerMTok": {
+                    "input": 0.15,
+                    "output": 0.5,
+                    "cachedInput": 0.03
+                },
+                "vision": true,
+                "status": "production",
+                "description": "ZhipuAI GLM 5.3 Flash — lighter/cheaper sibling of the default. Roster-listed only (GET /v1/models, 2026-09-28); not itself exercised in any code sample; fallback. vision:true and max_output_tokens are as declared by the roster response — the latter is null there, so it is omitted rather than invented."
+            },
+            "qwen-3.8-27b": {
+                "contextWindow": 262128,
+                "maxOutputTokens": 131072,
+                "pricingPerMTok": {
+                    "input": 0.45,
+                    "output": 3.2,
+                    "cachedInput": 0.25
+                },
+                "vision": true,
+                "status": "production",
+                "description": "Qwen 3.8 27B (Qwen/Qwen3.8-27B-FP8) — dense 27B VL model with native 262K context; the first entry in Neuralwatt's own /v1/models roster (GET, 2026-09-28) and one of the few catalog entries with a documented max_output_tokens ceiling (131,072) and image limits (max_images:20, max_images_total:200), so it is used as models.visionModel. Roster-listed and roster-priced only; fallback, not capability-probed (credential-free onboarding, no POST made)."
+            },
+            "qwen3.6-35b": {
+                "contextWindow": 262128,
+                "pricingPerMTok": {
+                    "input": 0.29,
+                    "output": 1.15,
+                    "cachedInput": 0.029
+                },
+                "vision": true,
+                "status": "production",
+                "description": "Qwen 3.6 35B — roster-listed only (GET /v1/models, 2026-09-28); not capability-probed; fallback. max_output_tokens is null in the roster response, so it is omitted here."
+            },
+            "kimi-k2.7-code": {
+                "contextWindow": 262128,
+                "pricingPerMTok": {
+                    "input": 0.95,
+                    "output": 4.0,
+                    "cachedInput": 0.095
+                },
+                "vision": true,
+                "status": "production",
+                "description": "MoonshotAI Kimi K2.7 Code (moonshotai/Kimi-K2.7-Code) — \"quant-agnostic canonical name for the Kimi K2.7 Code family\" per the roster's own description field (GET /v1/models, 2026-09-28). Roster-listed and roster-priced only; not capability-probed; fallback. max_output_tokens is null in the roster response, so it is omitted here."
+            },
+            "kimi-k3": {
+                "contextWindow": 1048560,
+                "pricingPerMTok": {
+                    "input": 3.0,
+                    "output": 15.0,
+                    "cachedInput": 0.3
+                },
+                "vision": true,
+                "status": "production",
+                "description": "MoonshotAI Kimi K3 — general-purpose sibling of the Code line, 1M context per the roster (GET /v1/models, 2026-09-28). Roster-listed only; not capability-probed; fallback. max_output_tokens is null in the roster response, so it is omitted here."
+            },
+            "deepseek-v4.1-flash": {
+                "contextWindow": 1048560,
+                "maxOutputTokens": 393216,
+                "pricingPerMTok": {
+                    "input": 0.15,
+                    "output": 0.6,
+                    "cachedInput": 0.015
+                },
+                "vision": true,
+                "status": "production",
+                "description": "DeepSeek V4.1 Flash — roster-listed with a documented max_output_tokens ceiling (393,216) and vision:true (GET /v1/models, 2026-09-28). Not capability-probed; fallback."
+            },
+            "gemma-4-31b": {
+                "contextWindow": 262128,
+                "maxOutputTokens": 16384,
+                "pricingPerMTok": {
+                    "input": 0.144,
+                    "output": 0.42,
+                    "cachedInput": 0.0144
+                },
+                "vision": true,
+                "status": "production",
+                "description": "NVIDIA's Gemma 4 31B build (owned_by/provider \"nvidia\" per the roster, GET /v1/models, 2026-09-28) — the only catalog entry from that provider label. Roster-listed with a documented max_output_tokens ceiling (16,384); not capability-probed; fallback."
+            }
+        },
+        "topModels": [
+            "glm-5.3",
+            "glm-5.3-flash",
+            "qwen-3.8-27b",
+            "qwen3.6-35b",
+            "kimi-k2.7-code",
+            "kimi-k3",
+            "deepseek-v4.1-flash",
+            "gemma-4-31b"
+        ]
+    },
+    "capabilities": {
+        "text": true,
+        "streaming": true,
+        "tools": "model-dependent",
+        "toolsWithStreaming": false,
+        "structuredOutput": false,
+        "structuredOutputWithTools": false,
+        "embeddings": false,
+        "thinking": true
+    },
+    "errorRules": [
+        {
+            "status": 401,
+            "pattern": "Invalid API key|invalid_api_key|authentication_error",
+            "class": "authentication",
+            "message": "Invalid Neuralwatt API key. Check {apiKeyEnvVar}. Get one at {setupUrl}"
+        },
+        {
+            "status": 404,
+            "pattern": "model_not_found|not\\s+found",
+            "class": "invalid-model",
+            "message": "Neuralwatt model '{model}' not found. Pick a current model from the unauthenticated GET https://api.neuralwatt.com/v1/models roster or {setupUrl}"
+        },
+        {
+            "status": 429,
+            "pattern": "rate_limit_exceeded|concurrent_budget_exceeded|tpm_uncached_exceeded",
+            "class": "rate-limit",
+            "message": "Neuralwatt rate limit hit for {model}. The error body includes retry_after and a retry_strategy — back off and retry rather than failing immediately."
+        }
+    ],
+    "setup": {
+        "url": "https://portal.neuralwatt.com",
+        "apiKeyFormat": null,
+        "billingPolicy": "free-with-card",
+        "instructions": [
+            "1. Visit: https://portal.neuralwatt.com and sign up",
+            "2. Add a payment method in Dashboard > Billing — new accounts start with a $0 balance, and adding a card grants $1.00 in free trial credit (no charge is made; the credit is valid for 30 days)",
+            "3. Navigate to Dashboard > API Keys and create a new key — copy it immediately, it is shown in full only once and starts with sk-",
+            "4. Set {apiKeyEnvVar} in your .env file"
+        ],
+        "description": "OpenAI-compatible endpoint at https://api.neuralwatt.com/v1 (\"Neuralwatt Cloud\"), serving 21 hosted open-weight models across Qwen, MoonshotAI Kimi, ZhipuAI GLM, DeepSeek and an NVIDIA Gemma build, with per-request energy and cost reporting (SSE comment lines / X-Request-Cost-USD headers). New accounts start at a $0 balance; adding a payment method grants a $1.00 trial credit valid 30 days, then usage is pay-per-token. Not all models support tool calling per Neuralwatt's own tool-calling guide — check a given model's capabilities on the /v1/models roster before relying on it."
+    },
+    "evidence": {
+        "rosterVerified": {
+            "date": "2026-09-28",
+            "method": "unauthenticated GET https://api.neuralwatt.com/v1/models — 21 models; credential-free, no key used",
+            "status": 200
+        },
+        "liveMatrix": null,
+        "addedInPR": "https://github.com/juspay/neurolink/pull/1848"
+    }
+}

package/dist/providers/catalog/ovhcloud.json

@@ -0,0 +1,232 @@
+{
+    "$schema": "./provider-catalog.schema.json",
+    "id": "ovhcloud",
+    "displayName": "OVHcloud AI Endpoints",
+    "aliases": [],
+    "tier": 2,
+    "wire": {
+        "baseURL": "https://oai.endpoints.kepler.ai.cloud.ovh.net/v1",
+        "envOverrides": {
+            "apiKey": "OVH_AI_ENDPOINTS_ACCESS_TOKEN"
+        }
+    },
+    "models": {
+        "default": "gpt-oss-120b",
+        "fallbacks": [
+            "gpt-oss-20b",
+            "Meta-Llama-3_3-70B-Instruct",
+            "Qwen3.6-27B",
+            "Mistral-Small-3.2-24B-Instruct-2506",
+            "Qwen3.5-9B"
+        ],
+        "defaultContextWindow": 131072,
+        "defaultMaxOutputTokens": 131072,
+        "catalog": {
+            "gpt-oss-120b": {
+                "contextWindow": 131072,
+                "maxOutputTokens": 131072,
+                "pricingPerMTok": {
+                    "input": 0.09,
+                    "output": 0.47
+                },
+                "vision": false,
+                "status": "production",
+                "description": "Recommended default — OpenAI's open-weight gpt-oss-120b, the exact model OVHcloud's own getting-started guide uses for every code sample (docs.ovhcloud.com/en/guides/public-cloud/ai-machine-learning/ai-endpoints-getting-started, retrieved 2026-09-28), and the most fully documented model on the public catalog card (ovhcloud.com/en/public-cloud/ai-endpoints/catalog/gpt-oss-120b/, retrieved 2026-09-28). That card's own \"Supported Features\" list names exactly Function calling, Reasoning and Streaming; its \"Output Formats\" list names raw_text, json_object and json_schema. The card's Python example builds the request with `Authorization: f\"Bearer {os.getenv('OVH_AI_ENDPOINTS_ACCESS_TOKEN')}\"` against https://oai.endpoints.kepler.ai.cloud.ovh.net/v1/chat/completions, and reads a `reasoning_effort` generation parameter (low/medium/high) plus a `reasoning_content` response field — the basis for capabilities.thinking. No sign-in or API key was used to read any of this; it is all public documentation plus the unauthenticated GET on /v1/models. contextWindow and maxOutputTokens are the live, unauthenticated /v1/models response's own context_length/max_completion_tokens fields for this id (GET run 2026-09-28, HTTP 200). Pricing is that same response's per-token USD pricing (prompt/completion fields) converted to $/MTok; the EUR-denominated marketing catalog card quotes the same model at €0.08 in / €0.4 out per Mtoken, consistent with the API's USD figures at roughly current EUR/USD. No live request of any kind (chat, streaming, tool call) was made against this model — the entry is docs- and roster-verified only; see evidence.liveMatrix (null)."
+            },
+            "gpt-oss-20b": {
+                "contextWindow": 131072,
+                "maxOutputTokens": 131072,
+                "pricingPerMTok": {
+                    "input": 0.05,
+                    "output": 0.18
+                },
+                "vision": false,
+                "status": "production",
+                "description": "OpenAI's smaller open-weight gpt-oss-20b. Listed on the public catalog card (ovhcloud.com/en/public-cloud/ai-endpoints/catalog/, retrieved 2026-09-28) as a \"Reasoning LLM\" with Function calling: Yes, priced €0.04 in / €0.15 out per Mtoken — consistent with this id's per-token USD pricing on the live, unauthenticated GET /v1/models response (2026-09-28, HTTP 200), which also supplies contextWindow/maxOutputTokens. Not individually probed; roster-verified fallback."
+            },
+            "Qwen3.6-27B": {
+                "contextWindow": 262144,
+                "maxOutputTokens": 262144,
+                "pricingPerMTok": {
+                    "input": 0.47,
+                    "output": 3.19
+                },
+                "vision": true,
+                "status": "production",
+                "description": "Qwen3.6-27B, listed on the public catalog card as \"Visual LLM (New)\" with Vision: Yes and Function calling: Yes (ovhcloud.com/en/public-cloud/ai-endpoints/catalog/, retrieved 2026-09-28), priced €0.4 in / €2.7 out per Mtoken there — consistent with this id's per-token USD pricing on the live, unauthenticated GET /v1/models (2026-09-28, HTTP 200), which also supplies contextWindow/maxOutputTokens. Not individually probed; roster-verified fallback. No image-input wire format is documented, so it is not set as models.visionModel."
+            },
+            "Qwen3.8-27B": {
+                "contextWindow": 262144,
+                "maxOutputTokens": 262144,
+                "pricingPerMTok": {
+                    "input": 0.47,
+                    "output": 3.19
+                },
+                "vision": true,
+                "status": "production",
+                "description": "Qwen3.8-27B, a newer sibling of Qwen3.6-27B on the same catalog card, same \"Visual LLM (New)\" row: Vision: Yes, Function calling: Yes, €0.4 in / €2.7 out per Mtoken (ovhcloud.com/en/public-cloud/ai-endpoints/catalog/, retrieved 2026-09-28) — consistent with this id's per-token USD pricing on the live GET /v1/models (2026-09-28). Not individually probed; roster-verified fallback."
+            },
+            "Qwen3.5-397B-A17B": {
+                "contextWindow": 262144,
+                "maxOutputTokens": 262144,
+                "pricingPerMTok": {
+                    "input": 0.71,
+                    "output": 4.25
+                },
+                "vision": true,
+                "status": "production",
+                "description": "Qwen3.5-397B-A17B, the largest model in this roster (397B-parameter MoE, ~17B active). Catalog card: \"Visual LLM (New)\", Vision: Yes, Function calling: Yes, €0.6 in / €3.6 out per Mtoken (ovhcloud.com/en/public-cloud/ai-endpoints/catalog/, retrieved 2026-09-28) — consistent with this id's per-token USD pricing on the live GET /v1/models (2026-09-28). Not individually probed; roster-verified fallback."
+            },
+            "Qwen3.5-9B": {
+                "contextWindow": 262144,
+                "maxOutputTokens": 262144,
+                "pricingPerMTok": {
+                    "input": 0.12,
+                    "output": 0.18
+                },
+                "vision": true,
+                "status": "production",
+                "description": "Qwen3.5-9B, the smallest of the Qwen3.x \"Visual LLM (New)\" row: Vision: Yes, Function calling: Yes, €0.1 in / €0.15 out per Mtoken (ovhcloud.com/en/public-cloud/ai-endpoints/catalog/, retrieved 2026-09-28) — consistent with this id's per-token USD pricing on the live GET /v1/models (2026-09-28). Not individually probed; roster-verified fallback."
+            },
+            "Meta-Llama-3_3-70B-Instruct": {
+                "contextWindow": 131072,
+                "maxOutputTokens": 131072,
+                "pricingPerMTok": {
+                    "input": 0.74,
+                    "output": 0.74
+                },
+                "vision": false,
+                "status": "production",
+                "description": "Meta Llama 3.3 70B Instruct, a plain non-reasoning chat model. Catalog card: \"LLM\" category, Vision: No, Function calling: Yes, €0.67 in / €0.67 out per Mtoken (ovhcloud.com/en/public-cloud/ai-endpoints/catalog/, retrieved 2026-09-28) — consistent with this id's per-token USD pricing on the live GET /v1/models (2026-09-28). Not individually probed; roster-verified fallback."
+            },
+            "Qwen2.5-VL-72B-Instruct": {
+                "contextWindow": 32768,
+                "maxOutputTokens": 32768,
+                "pricingPerMTok": {
+                    "input": 1.01,
+                    "output": 1.01
+                },
+                "vision": true,
+                "status": "production",
+                "description": "Qwen2.5-VL-72B-Instruct — a dedicated vision-language model (\"VL\"). Catalog card: \"Visual LLM\" category, Vision: Yes, Function calling: No, €0.91 in / €0.91 out per Mtoken (ovhcloud.com/en/public-cloud/ai-endpoints/catalog/, retrieved 2026-09-28) — the only model in this roster the vendor's own catalog explicitly marks Vision: Yes together with a specific, non-reasoning-family name, so it is set as models.visionModel. No image wire format (base64 vs. URL, content-parts shape) is documented, and no image request was sent (credential-free). Pricing/context/output-cap consistent with the live GET /v1/models (2026-09-28). Not individually probed; roster-verified."
+            },
+            "Qwen3-Coder-30B-A3B-Instruct": {
+                "contextWindow": 262144,
+                "maxOutputTokens": 262144,
+                "pricingPerMTok": {
+                    "input": 0.07,
+                    "output": 0.26
+                },
+                "vision": false,
+                "status": "production",
+                "description": "Qwen3-Coder-30B-A3B-Instruct — present on the live, unauthenticated GET /v1/models response (2026-09-28, HTTP 200: id, context_length=262144, max_completion_tokens=262144, USD pricing) but not yet on the public marketing catalog card (ovhcloud.com/en/public-cloud/ai-endpoints/catalog/, checked 2026-09-28 — that page lists 20 of the API's 24 live ids; this is one of the 4 not yet listed there). vision:false follows the model family naming (a code-instruct Qwen3 build, no \"VL\"/visual designation) rather than an explicit per-model vision flag; tools support is likewise not individually confirmed for this id — roster-verified only, not capability-probed."
+            },
+            "Mistral-Small-3.2-24B-Instruct-2506": {
+                "contextWindow": 131072,
+                "maxOutputTokens": 131072,
+                "pricingPerMTok": {
+                    "input": 0.1,
+                    "output": 0.31
+                },
+                "vision": false,
+                "status": "production",
+                "description": "Mistral Small 3.2 24B Instruct (2506 snapshot) — present on the live, unauthenticated GET /v1/models response (2026-09-28, HTTP 200) but not yet on the public marketing catalog card (checked 2026-09-28, one of the 4 live ids not yet listed there). Mistral's own upstream Small 3.2 release is text-only in its OVH deployment name (no \"VL\"/visual designation); vision:false and tools support are not individually confirmed for this id — roster-verified only."
+            },
+            "Mistral-7B-Instruct-v0.3": {
+                "contextWindow": 65536,
+                "maxOutputTokens": 65536,
+                "pricingPerMTok": {
+                    "input": 0.11,
+                    "output": 0.11
+                },
+                "vision": false,
+                "status": "production",
+                "description": "Mistral 7B Instruct v0.3 — present on the live, unauthenticated GET /v1/models response (2026-09-28, HTTP 200) but not on the public marketing catalog card (checked 2026-09-28). Small, plain instruct model; vision:false and tools support are not individually confirmed for this id — roster-verified fallback only."
+            },
+            "Mistral-Nemo-Instruct-2407": {
+                "contextWindow": 65536,
+                "maxOutputTokens": 65536,
+                "pricingPerMTok": {
+                    "input": 0.14,
+                    "output": 0.14
+                },
+                "vision": false,
+                "status": "production",
+                "description": "Mistral Nemo Instruct (2407 snapshot) — present on the live, unauthenticated GET /v1/models response (2026-09-28, HTTP 200) but not on the public marketing catalog card (checked 2026-09-28). Plain instruct model; vision:false and tools support are not individually confirmed for this id — roster-verified fallback only."
+            },
+            "Qwen3Guard-Gen-8B": {
+                "contextWindow": 32768,
+                "maxOutputTokens": 16384,
+                "pricingPerMTok": {
+                    "input": 0,
+                    "output": 0
+                },
+                "vision": false,
+                "status": "preview",
+                "description": "Qwen3Guard-Gen-8B — a moderation/safety classifier, not a general-purpose chat model. Catalog card: \"LLM Guard (Beta)\" category, priced Free (ovhcloud.com/en/public-cloud/ai-endpoints/catalog/, retrieved 2026-09-28) — the live GET /v1/models response (2026-09-28) shows prompt/completion pricing of 0 for this id, confirming it. status is \"preview\" because the vendor explicitly labels it Beta. Kept out of models.fallbacks/topModels since it answers moderation prompts, not open-ended generation."
+            },
+            "Qwen3Guard-Gen-0.6B": {
+                "contextWindow": 32768,
+                "maxOutputTokens": 16384,
+                "pricingPerMTok": {
+                    "input": 0,
+                    "output": 0
+                },
+                "vision": false,
+                "status": "preview",
+                "description": "Qwen3Guard-Gen-0.6B — smaller sibling of Qwen3Guard-Gen-8B, same moderation/safety-classifier role. Catalog card: \"LLM Guard (Beta)\" category, priced Free (ovhcloud.com/en/public-cloud/ai-endpoints/catalog/, retrieved 2026-09-28); the live GET /v1/models response (2026-09-28) confirms 0 pricing for this id. status \"preview\" per the vendor's own Beta label. Kept out of models.fallbacks/topModels for the same reason as the 8B variant."
+            }
+        },
+        "topModels": [
+            "gpt-oss-120b",
+            "gpt-oss-20b",
+            "Qwen3.6-27B",
+            "Qwen3.8-27B",
+            "Qwen3.5-397B-A17B",
+            "Qwen3.5-9B",
+            "Meta-Llama-3_3-70B-Instruct",
+            "Qwen2.5-VL-72B-Instruct",
+            "Mistral-Small-3.2-24B-Instruct-2506"
+        ],
+        "visionModel": "Qwen2.5-VL-72B-Instruct"
+    },
+    "capabilities": {
+        "text": true,
+        "streaming": true,
+        "tools": "model-dependent",
+        "toolsWithStreaming": false,
+        "structuredOutput": true,
+        "structuredOutputWithTools": false,
+        "embeddings": false,
+        "thinking": true
+    },
+    "errorRules": [
+        {
+            "status": 403,
+            "pattern": "Authentication Failed",
+            "class": "authentication",
+            "message": "Invalid or revoked OVHcloud AI Endpoints access token. Check {apiKeyEnvVar}. Get one at {setupUrl}"
+        }
+    ],
+    "setup": {
+        "url": "https://www.ovhcloud.com/en/public-cloud/ai-endpoints/",
+        "apiKeyFormat": null,
+        "billingPolicy": "free-with-card",
+        "instructions": [
+            "1. Visit: https://www.ovhcloud.com/en/public-cloud/ai-endpoints/ and click Start now, which signs you in to (or creates) an OVHcloud Public Cloud manager account",
+            "2. In the manager, open Public Cloud > your project > AI & Machine Learning > AI Endpoints, and add a payment method to the project — OVHcloud states that projects in Discovery mode (no payment method) cannot use the service",
+            "3. Create an AI Endpoints API access token in that section; it is shown only once, so store it yourself, and note it can be given an expiration date and revoked",
+            "4. Set {apiKeyEnvVar} in your .env file (the vendor's own docs name this exact variable)"
+        ],
+        "description": "OpenAI-compatible endpoint at https://oai.endpoints.kepler.ai.cloud.ovh.net/v1, documented as \"compatible with the OpenAI specifications\" (docs.ovhcloud.com/en/guides/public-cloud/ai-machine-learning/ai-endpoints-getting-started). Several models (both Qwen3Guard classifiers, stable-diffusion-xl, the four NVIDIA Riva TTS voices) are billed at €0/free, but the account itself requires a payment method on the Public Cloud project before any API key can be used — OVHcloud's own docs: \"projects in Discovery mode (without a payment method) cannot use the service.\" A very limited anonymous tier also exists (2 requests/minute per IP per model, undocumented auth shape) but is not what {apiKeyEnvVar} configures. Authenticated rate limit is documented as 400 requests/minute per project per model; over that returns HTTP 429. This catalog entry covers only the chat-completions models; OVHcloud AI Endpoints also serves speech-to-text (Whisper), text-to-speech (NVIDIA Riva), image generation (Stable Diffusion XL) and embeddings (bge-m3, bge-multilingual-gemma2, Qwen3-Embedding-8B) models on the same platform, which are out of scope for this Tier-2 chat-completions integration."
+    },
+    "evidence": {
+        "rosterVerified": {
+            "date": "2026-09-28",
+            "method": "unauthenticated GET https://oai.endpoints.kepler.ai.cloud.ovh.net/v1/models — 24 models; credential-free, no key used",
+            "status": 200
+        },
+        "liveMatrix": null,
+        "addedInPR": "https://github.com/juspay/neurolink/pull/1848"
+    }
+}

package/dist/providers/catalog/pareto-inference.json

@@ -0,0 +1,74 @@
+{
+    "$schema": "./provider-catalog.schema.json",
+    "id": "pareto-inference",
+    "displayName": "Pareto Inference",
+    "aliases": [],
+    "tier": 2,
+    "wire": {
+        "baseURL": "https://api.paretoinference.com/v1"
+    },
+    "models": {
+        "default": "z-ai/glm-5.3-flash",
+        "fallbacks": ["z-ai/glm-5.3-flash"],
+        "defaultContextWindow": 131072,
+        "defaultMaxOutputTokens": 131072,
+        "catalog": {
+            "z-ai/glm-5.3-flash": {
+                "maxOutputTokens": 131072,
+                "pricingPerMTok": {
+                    "input": 0.03,
+                    "output": 0.1,
+                    "cachedInput": 0.006
+                },
+                "vision": false,
+                "status": "production",
+                "description": "Recommended - the only model Pareto serves; GLM 5.3 Flash (Zhipu/Z.ai), run on Pareto's own GPUs. Confirmed live via an unauthenticated GET https://api.paretoinference.com/v1/models on 2026-09-28 (HTTP 200, exactly one id: z-ai/glm-5.3-flash, owned_by z-ai) - no key used, matching the docs' note that GET /models needs no API key. The docs also document a bare alias, glm-5.3-flash, for the same model (https://docs.paretoinference.com/api-reference, retrieved 2026-09-28). Wire is OpenAI Chat Completions: POST /chat/completions with required model + messages ('Use the OpenAI message format'), the answer read from choices[0].message.content, SSE streaming with text in choices[0].delta.content and tool calls in choices[0].delta.tool_calls, a final usage event when stream_options.include_usage is true, and a terminal 'data: [DONE]' (https://docs.paretoinference.com/api-reference, retrieved 2026-09-28). Output tokens (max_tokens / max_completion_tokens - send only one) range 1-131,072, defaulting to 131,072 when both are omitted - that figure is this entry's defaultMaxOutputTokens and maxOutputTokens. Pareto documents no separate context-length figure anywhere: 'We do not disclose the technical details of our GLM 5.3 Flash serving system' (https://docs.paretoinference.com/how-pareto-works, retrieved 2026-09-28), and no context-window number appears on the quickstart, API reference, how-it-works, errors, or pricing pages, so models.defaultContextWindow above (131,072) reuses the documented output ceiling as a conservative floor rather than a vendor-stated context length - correct it once Pareto publishes one or an authenticated probe reveals one; contextWindow is deliberately left unset on this model entry rather than restating that same undocumented floor as a per-model fact. Pricing from https://docs.paretoinference.com/pricing, retrieved 2026-09-28: $0.03 per 1M input tokens, $0.10 per 1M output tokens, $0.006 per 1M cached-input tokens ('a cache hit is not guaranteed'). tools ('Sends OpenAI function tools') and response_format ('Requests a structured response') are both documented request parameters, but the same request-parameters table hedges both of them - plus tool_choice and logprobs - with 'Model support can differ'; there is no model-specific confirmation for z-ai/glm-5.3-flash and no live probe was possible credential-free, so capabilities.tools is 'model-dependent' here and toolsWithStreaming/structuredOutput/structuredOutputWithTools are left false rather than asserted from a hedged parameter table. reasoning_effort ('Sets the reasoning level') is documented as a plain, unhedged request parameter - unlike tools/response_format/tool_choice/logprobs it carries no 'Model support can differ' caveat - which is the basis for capabilities.thinking: true. Vision/image input and an embeddings endpoint are not documented anywhere on docs.paretoinference.com, so vision stays false and models.visionModel is left unset. No live capability probe was run: this is credential-free onboarding, no Pareto API key was ever created, and no POST request was ever sent to the vendor - see evidence.liveMatrix: null."
+            }
+        },
+        "topModels": ["z-ai/glm-5.3-flash"]
+    },
+    "capabilities": {
+        "text": true,
+        "streaming": true,
+        "tools": "model-dependent",
+        "toolsWithStreaming": false,
+        "structuredOutput": false,
+        "structuredOutputWithTools": false,
+        "embeddings": false,
+        "thinking": true
+    },
+    "errorRules": [
+        {
+            "status": 401,
+            "class": "authentication",
+            "message": "Invalid Pareto Inference API key. Check {apiKeyEnvVar}. Pareto documents 401 as \"the chat API key is missing, invalid, or no longer active\" (unauthenticated GET /models needs no key, so this only affects POST /chat/completions) but publishes no example error body - get or replace a key at {setupUrl}"
+        },
+        {
+            "status": 429,
+            "pattern": "credit_exhausted|credit_insufficient|model_capacity",
+            "class": "rate-limit",
+            "message": "Pareto rate/credit limit hit for {model}. Documented causes: credit_exhausted (no credits left - buy more), credit_insufficient (held credits below what max_tokens could use - lower max_tokens or buy credits), or model_capacity (the model is at capacity, not an account limit - honor Retry-After). See {setupUrl}"
+        }
+    ],
+    "setup": {
+        "url": "https://paretoinference.com/dashboard",
+        "apiKeyFormat": null,
+        "billingPolicy": "no-free-tier",
+        "instructions": [
+            "1. Visit: https://paretoinference.com/dashboard and sign in",
+            "2. Add a payment card - Pareto has no free tier; a card and prepaid credits are required before a key is issued (https://docs.paretoinference.com/pricing)",
+            "3. Buy prepaid credits, then create your API key from the dashboard",
+            "4. Set {apiKeyEnvVar} in your .env file"
+        ],
+        "description": "OpenAI-compatible Chat Completions endpoint at https://api.paretoinference.com/v1, serving exactly one hosted model - z-ai/glm-5.3-flash (GLM 5.3 Flash), also reachable via the documented alias glm-5.3-flash - on Pareto's own GPUs. No free tier: a payment card and prepaid credits are required before a key is issued. Pareto offers zero data retention (ZDR) and does not keep request content after a request finishes."
+    },
+    "evidence": {
+        "rosterVerified": {
+            "date": "2026-09-28",
+            "method": "unauthenticated GET https://api.paretoinference.com/v1/models - 1 model; credential-free, no key used",
+            "status": 200
+        },
+        "liveMatrix": null,
+        "addedInPR": "https://github.com/juspay/neurolink/pull/1848"
+    }
+}

package/dist/providers/catalog/sarvam.json

@@ -0,0 +1,72 @@
+{
+    "$schema": "./provider-catalog.schema.json",
+    "id": "sarvam",
+    "displayName": "Sarvam AI",
+    "aliases": [],
+    "tier": 2,
+    "wire": {
+        "baseURL": "https://api.sarvam.ai/v1"
+    },
+    "models": {
+        "default": "sarvam-105b",
+        "fallbacks": ["sarvam-105b-conversations"],
+        "defaultContextWindow": 128000,
+        "defaultMaxOutputTokens": 2048,
+        "catalog": {
+            "sarvam-105b": {
+                "contextWindow": 128000,
+                "vision": false,
+                "status": "production",
+                "description": "Recommended - Sarvam's flagship chat model, documented as \"(128K context) for complex reasoning and agentic tasks\" (https://docs.sarvam.ai/api-reference-docs/chat/chat-completions, retrieved 2026-09-28). The /v1/chat/completions endpoint serves only sarvam-105b and sarvam-105b-conversations per the vendor's own reference — confirmed live by an unauthenticated GET https://api.sarvam.ai/v1/models on 2026-09-28 (HTTP 200, exactly these 2 ids, no key used). Request/response wire shape is OpenAI-compatible: messages[] with system/developer/user/assistant/tool roles, choices[].message/finish_reason (stop, length, tool_calls, content_filter, deprecated function_call), object: \"chat.completion\". tools + tool_choice (function calling) and response_format json_schema/json_object (Structured Outputs) are both documented generically for this endpoint, not per model — none of this was live-probed, since onboarding was credential-free (no API key, no POST request made to the vendor's API). Sarvam-specific reasoning_effort (low/high/max, default medium, disable by sending null) plus a reasoning_content field on the response message are documented — the basis for capabilities.thinking here. max_tokens defaults to 2048 if the caller omits it and the docs state no maximum output ceiling, so defaultMaxOutputTokens below reflects that documented default rather than a hard cap; maxOutputTokens is left unset on this entry rather than inventing one. Pricing is documented only in Indian Rupees — ₹29.28 input / ₹73.2 output / ₹10.98 cached-input per 1M tokens (https://docs.sarvam.ai/api/getting-started/pricing.md, retrieved 2026-09-28); pricingPerMTok is left unset here rather than placing an INR figure in a field every other catalog entry uses for USD, since the vendor states no INR/USD rate to convert by without inventing a number. The message content-part schema also declares an image_url part (standard OpenAI-compatible shape), but the vendor documents no vision/multimodal capability for either model, so vision stays false and models.visionModel is left unset.",
+                "enumMember": "SARVAM_105B"
+            },
+            "sarvam-105b-conversations": {
+                "contextWindow": 32000,
+                "vision": false,
+                "status": "production",
+                "description": "Sarvam's conversational variant, documented as \"(32K context) for real-time conversational and voice-agent workloads\" (https://docs.sarvam.ai/api-reference-docs/chat/chat-completions, retrieved 2026-09-28) — kept as the fallback behind the larger-context sarvam-105b. Same endpoint, same OpenAI-compatible wire shape, tools/response_format documentation and reasoning_effort/reasoning_content fields as sarvam-105b (see that entry); confirmed live alongside it by the same unauthenticated GET https://api.sarvam.ai/v1/models on 2026-09-28 (HTTP 200). Pricing per the vendor's pricing page is identical to sarvam-105b — ₹29.28 input / ₹73.2 output / ₹10.98 cached-input per 1M tokens, stated only in INR — so pricingPerMTok is left unset for the same reason.",
+                "enumMember": "SARVAM_105B_CONVERSATIONS"
+            }
+        },
+        "topModels": ["sarvam-105b", "sarvam-105b-conversations"]
+    },
+    "capabilities": {
+        "text": true,
+        "streaming": true,
+        "tools": true,
+        "toolsWithStreaming": true,
+        "structuredOutput": true,
+        "structuredOutputWithTools": false,
+        "embeddings": false,
+        "thinking": true
+    },
+    "errorRules": [
+        {
+            "status": 403,
+            "pattern": "invalid_api_key_error|Invalid API key",
+            "class": "authentication",
+            "message": "Invalid Sarvam AI API key. Check {apiKeyEnvVar}. Get one at {setupUrl}"
+        }
+    ],
+    "setup": {
+        "url": "https://dashboard.sarvam.ai",
+        "apiKeyFormat": "^sk_[A-Za-z0-9_-]+$",
+        "billingPolicy": "free-tier",
+        "instructions": [
+            "1. Visit: https://dashboard.sarvam.ai and sign in or create an account",
+            "2. Create an API subscription key from the dashboard (documented format: sk_xxx)",
+            "3. Every new account receives ₹100 in starter credits, then usage-based billing applies (https://docs.sarvam.ai/api/getting-started/pricing.md) — no ongoing free tier beyond that starter credit is documented",
+            "4. Set {apiKeyEnvVar} in your .env file"
+        ],
+        "description": "OpenAI-compatible endpoint at https://api.sarvam.ai/v1, serving only sarvam-105b and sarvam-105b-conversations via POST /v1/chat/completions. The vendor's own API reference authenticates its raw HTTP examples with a standard Authorization: Bearer <token> header — what NeuroLink's generic OpenAI-compatible provider sends. Sarvam's stated preference is instead an api-subscription-key header, which its Python/TypeScript SDKs use; both are documented as accepted on every endpoint, so no custom auth-header quirk is needed here."
+    },
+    "evidence": {
+        "rosterVerified": {
+            "date": "2026-09-28",
+            "method": "unauthenticated GET https://api.sarvam.ai/v1/models — 2 models (sarvam-105b, sarvam-105b-conversations); credential-free, no key used",
+            "status": 200
+        },
+        "liveMatrix": null,
+        "addedInPR": "https://github.com/juspay/neurolink/pull/1848"
+    }
+}

package/dist/providers/catalog/schema.d.ts

@@ -48,8 +48,8 @@ export declare const providerCatalogJsonSchema: z.ZodObject<{
             vision: z.ZodBoolean;
             status: z.ZodEnum<{
                 production: "production";
-                retired: "retired";
                 preview: "preview";
+                retired: "retired";
             }>;
             description: z.ZodString;
             enumMember: z.ZodOptional<z.ZodString>;
@@ -96,9 +96,9 @@ export declare const providerCatalogJsonSchema: z.ZodObject<{
         url: z.ZodString;
         apiKeyFormat: z.ZodNullable<z.ZodString>;
         billingPolicy: z.ZodEnum<{
+            "no-free-tier": "no-free-tier";
             "free-tier": "free-tier";
             "free-with-card": "free-with-card";
-            "no-free-tier": "no-free-tier";
         }>;
         instructions: z.ZodArray<z.ZodString>;
         description: z.ZodOptional<z.ZodString>;