@neon/ai-sdk-provider 0.8.0 → 0.10.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/README.md +21 -9
- package/dist/lib/neon-capabilities.d.ts.map +1 -1
- package/dist/lib/neon-capabilities.js +4 -4
- package/dist/lib/neon-capabilities.js.map +1 -1
- package/dist/lib/neon-chat-options.d.ts +2 -2
- package/dist/lib/neon-chat-options.d.ts.map +1 -1
- package/dist/lib/neon-chat-options.js +9 -2
- package/dist/lib/neon-chat-options.js.map +1 -1
- package/dist/lib/neon-model-capabilities.d.ts.map +1 -1
- package/dist/lib/neon-model-capabilities.js +51 -7
- package/dist/lib/neon-model-capabilities.js.map +1 -1
- package/dist/package.js +1 -1
- package/package.json +3 -2
package/README.md
CHANGED
|
@@ -4,7 +4,7 @@ Community [Vercel AI SDK](https://ai-sdk.dev) provider for the [Neon](https://ne
|
|
|
4
4
|
|
|
5
5
|
The Neon AI Gateway is **branch-scoped**: each Neon project branch gets its own gateway host, and a platform token authorizes requests for that branch. Use the same `neon(modelId)` API across the branch's model catalog; the provider selects Anthropic Messages, OpenAI Responses, or Chat Completions for each model.
|
|
6
6
|
|
|
7
|
-
Use canonical model ids such as `gpt-5-mini`, `llama-4-maverick`, and `gemini-3-flash`, matching
|
|
7
|
+
Use canonical model ids such as `gpt-5-mini`, `llama-4-maverick`, and `gemini-3-flash`, matching [neon.com/models](https://neon.com/models). The typed catalog includes the known model ids, and arbitrary strings are accepted so newly available models work before the types update.
|
|
8
8
|
|
|
9
9
|
## Install
|
|
10
10
|
|
|
@@ -69,17 +69,20 @@ Which ids your branch serves is account-specific during the beta; `GET $NEON_AI_
|
|
|
69
69
|
|
|
70
70
|
Upstream backends behind the gateway accept different subsets of the OpenAI-style parameters the AI SDK emits, and sending one an upstream rejects is a hard `400`. The provider drops those before the request and records a warning in `result.warnings`, so a call succeeds instead of failing on a parameter you passed in good faith.
|
|
71
71
|
|
|
72
|
-
These rules apply on the Anthropic Messages and Chat Completions routes. The Responses route (every `gpt-*` id) relies on the upstream OpenAI model's own stripping instead, so its behaviour
|
|
72
|
+
These rules apply on the Anthropic Messages and Chat Completions routes. The Responses route (every `gpt-*` id) relies on the upstream OpenAI model's own stripping instead, so its behaviour and its warning wording are the AI SDK's rather than described here — but `getNeonModelCapabilities` still answers for those ids, and the `gpt-5` row below is what it reports rather than something this provider drops.
|
|
73
73
|
|
|
74
74
|
| Models | Dropped |
|
|
75
75
|
| --- | --- |
|
|
76
|
-
| Claude 4.7 and newer (`claude-opus-4-7`, `claude-opus-4-8`, `claude-opus-5`, `claude-sonnet-5`, `claude-fable-5`) | `temperature`, `topP` |
|
|
76
|
+
| Claude 4.7 and newer (`claude-opus-4-7`, `claude-opus-4-8`, `claude-opus-5`, `claude-sonnet-5`, `claude-fable-5`, `claude-fable-5-1`) | `temperature`, `topP` |
|
|
77
77
|
| All Claude | `frequencyPenalty`, `presencePenalty`, `seed`, `reasoningEffort`, and `topP` when you set both `temperature` and `topP` (Anthropic accepts only one) |
|
|
78
78
|
| `gpt-oss` | `frequencyPenalty`, `presencePenalty`, `seed`, `stopSequences` |
|
|
79
79
|
| Qwen, Gemma | `frequencyPenalty`, `presencePenalty`, `seed` |
|
|
80
|
-
| `glm-5-2`, `inkling` | `frequencyPenalty`, `presencePenalty` |
|
|
80
|
+
| `glm-5-2`, `glm-5-3-flash`, `inkling`, `kimi-k3` | `frequencyPenalty`, `presencePenalty` |
|
|
81
81
|
| Meta Llama | `frequencyPenalty`, `presencePenalty`, `seed` |
|
|
82
|
-
|
|
|
82
|
+
| `gemini-3-5-flash-lite` | `frequencyPenalty`, `presencePenalty` |
|
|
83
|
+
| `gemini-3-6-flash` | `temperature`, `topP`, `frequencyPenalty`, `presencePenalty` |
|
|
84
|
+
| Every other Gemini | nothing |
|
|
85
|
+
| `gpt-5`, `gpt-5-mini`, `gpt-5-nano`, `gpt-5-5-pro` (reported only — Responses strips these upstream) | `temperature`, `topP` |
|
|
83
86
|
|
|
84
87
|
**`temperature` on a Claude 5 model therefore has no effect.** That is the gateway's behaviour, not a provider choice; sending it returns `does not support the temperature parameter`. Steer those models with Anthropic's own effort control instead — note that this is `effort`, not the OpenAI-style `reasoningEffort`, which every Claude id drops:
|
|
85
88
|
|
|
@@ -97,10 +100,13 @@ Query the rules directly rather than hardcoding them:
|
|
|
97
100
|
import { getNeonModelCapabilities } from '@neon/ai-sdk-provider';
|
|
98
101
|
|
|
99
102
|
getNeonModelCapabilities('claude-opus-5').supportsTemperature; // false
|
|
103
|
+
getNeonModelCapabilities('gpt-6-astra').supportsTemperature; // false
|
|
100
104
|
getNeonModelCapabilities('glm-5-2').supportsPenalties; // false
|
|
101
105
|
```
|
|
102
106
|
|
|
103
|
-
|
|
107
|
+
**The rules are per model, not per family.** Two Gemini ids reject penalties while their siblings accept them, and `gemini-3-6-flash` rejects `temperature` and `topP` outright, so reading a row for "Gemini" is not enough — check the id. Every dropped-options table row was measured against the gateway rather than inherited from the upstream provider's own documentation, which disagrees in both directions. `getNeonModelCapabilities('gpt-6-astra').supportsTemperature` is `false` because [neon.com/models.json](https://neon.com/models.json) lists `temperature: false` for that id; it is not a live gateway measurement.
|
|
108
|
+
|
|
109
|
+
A model none of these rules match is left untouched, so a brand-new id gets the gateway's own error rather than a guess. That is also why the table can lag: an id added since the last release inherits the permissive default until someone measures it.
|
|
104
110
|
|
|
105
111
|
## Image generation
|
|
106
112
|
|
|
@@ -211,11 +217,17 @@ So `store` is only refused on the Responses route; the same option is ignored el
|
|
|
211
217
|
|
|
212
218
|
## End-to-end tests
|
|
213
219
|
|
|
214
|
-
|
|
220
|
+
These run against a real gateway, so they need credentials — one of two ways:
|
|
215
221
|
|
|
216
222
|
```bash
|
|
217
|
-
cp .env.example .env
|
|
223
|
+
cp .env.example .env
|
|
224
|
+
# Either NEON_API_KEY for a throwaway org, and the suite creates a project, mints a
|
|
225
|
+
# credential scoped to `ai_gateway:invoke`, and deletes both afterwards.
|
|
226
|
+
# Or NEON_AI_GATEWAY_BASE_URL + NEON_AI_GATEWAY_TOKEN from `neon env pull`, to run
|
|
227
|
+
# against a branch you already have.
|
|
218
228
|
pnpm test:e2e
|
|
219
229
|
```
|
|
220
230
|
|
|
221
|
-
|
|
231
|
+
A run with neither **fails**; it does not skip. The gateway is on every branch and needs no provisioning, so the API-key path is the one CI uses — no gateway token is stored as a secret.
|
|
232
|
+
|
|
233
|
+
The matrix covers one published catalog model per family (Anthropic, OpenAI, Codex, Gemini, Meta, Alibaba, Zhipu, Thinking Machines) across `generateText`, `streamText`, `generateObject`, tool calling, and `neon.tools.imageGeneration`. `generateObject` and tool calling run on the subset of families where they are verified (see [Capabilities](#capabilities)). A family whose representative id the branch does not serve skips its cases, and a single assertion fails the run listing exactly which pinned ids went missing — so a shrinking catalog is reported rather than silently reducing coverage. Model access is granted per account, so the account behind the key needs every id in `MATRIX_MODELS`. The suite also fetches the live `/v1/models` catalog and calls every currently enabled model with both AI SDK 6 and AI SDK 7.
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"neon-capabilities.d.ts","names":[],"sources":["../../src/lib/neon-capabilities.ts"],"mappings":";;;;;;AAoBA;AAAqC;AAE3B;AACI,iBAHE,qBAAA,CAGF,OAAA,EAAA,MAAA,EAAA,OAAA,EADJ,0BACI,CAAA,EAAA;EAAsC,OAAA,EAAtC,0BAAsC;EAAe,QAAA,EAAf,eAAe,EAAA;
|
|
1
|
+
{"version":3,"file":"neon-capabilities.d.ts","names":[],"sources":["../../src/lib/neon-capabilities.ts"],"mappings":";;;;;;AAoBA;AAAqC;AAE3B;AACI,iBAHE,qBAAA,CAGF,OAAA,EAAA,MAAA,EAAA,OAAA,EADJ,0BACI,CAAA,EAAA;EAAsC,OAAA,EAAtC,0BAAsC;EAAe,QAAA,EAAf,eAAe,EAAA;AA6GnE,CAAA;AAAwC;AAAQ;AAAiB;AAAA,iBAAjD,wBAAA,CAAiD,KAAA,EAAjB,eAAiB,EAAA,CAAA,EAAA,eAAA,CAAA,yBAAA,EAAA,yBAAA,CAAA"}
|
|
@@ -1,8 +1,8 @@
|
|
|
1
1
|
import { getNeonModelCapabilities } from "./neon-model-capabilities.js";
|
|
2
2
|
//#region src/lib/neon-capabilities.ts
|
|
3
|
-
const PENALTY_DETAILS = "
|
|
4
|
-
const GENERIC_DETAILS = "The
|
|
5
|
-
const REASONING_EFFORT_DETAILS = "
|
|
3
|
+
const PENALTY_DETAILS = "The request was sent without it. Check getNeonModelCapabilities('<model-id>').supportsPenalties before setting penalties.";
|
|
4
|
+
const GENERIC_DETAILS = "The request was sent without it. Call getNeonModelCapabilities('<model-id>') to see which sampling parameters this model takes, or steer it through the prompt.";
|
|
5
|
+
const REASONING_EFFORT_DETAILS = "The request was sent without it. Claude takes providerOptions.anthropic.effort instead.";
|
|
6
6
|
/**
|
|
7
7
|
* Drop call options the resolved model's upstream backend does not accept and
|
|
8
8
|
* collect a warning for each, so callers get a clear signal instead of a hard
|
|
@@ -20,7 +20,7 @@ function applyNeonCapabilities(modelId, options) {
|
|
|
20
20
|
});
|
|
21
21
|
};
|
|
22
22
|
const unrecognizedClaude = caps.claudeSamplingUnrecognized === true;
|
|
23
|
-
const samplingDetails = unrecognizedClaude ? "This Claude version was not recognised. Claude 4.7 and newer reject sampling parameters, so
|
|
23
|
+
const samplingDetails = unrecognizedClaude ? "This Claude version was not recognised. Claude 4.7 and newer reject sampling parameters, so the request was sent without it as a precaution; set `providerOptions.anthropic.effort` instead." : caps.family === "anthropic" ? "The request was sent without it. Claude 4.7 and newer reject sampling parameters; use `providerOptions.anthropic.effort` (low | medium | high | xhigh | max) to steer these models." : GENERIC_DETAILS;
|
|
24
24
|
const samplingType = unrecognizedClaude ? "compatibility" : "unsupported";
|
|
25
25
|
if (options.temperature != null && !caps.supportsTemperature) {
|
|
26
26
|
patch.temperature = void 0;
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"neon-capabilities.js","names":[],"sources":["../../src/lib/neon-capabilities.ts"],"sourcesContent":["import type {\n\tLanguageModelV3CallOptions,\n\tLanguageModelV3StreamPart,\n\tSharedV3ProviderOptions,\n\tSharedV3Warning,\n} from \"@ai-sdk/provider\";\nimport { getNeonModelCapabilities } from \"./neon-model-capabilities.js\";\n\nconst PENALTY_DETAILS =\n\t\"
|
|
1
|
+
{"version":3,"file":"neon-capabilities.js","names":[],"sources":["../../src/lib/neon-capabilities.ts"],"sourcesContent":["import type {\n\tLanguageModelV3CallOptions,\n\tLanguageModelV3StreamPart,\n\tSharedV3ProviderOptions,\n\tSharedV3Warning,\n} from \"@ai-sdk/provider\";\nimport { getNeonModelCapabilities } from \"./neon-model-capabilities.js\";\n\nconst PENALTY_DETAILS =\n\t\"The request was sent without it. Check getNeonModelCapabilities('<model-id>').supportsPenalties before setting penalties.\";\nconst GENERIC_DETAILS =\n\t\"The request was sent without it. Call getNeonModelCapabilities('<model-id>') to see which sampling parameters this model takes, or steer it through the prompt.\";\nconst REASONING_EFFORT_DETAILS =\n\t\"The request was sent without it. Claude takes providerOptions.anthropic.effort instead.\";\n\n/**\n * Drop call options the resolved model's upstream backend does not accept and\n * collect a warning for each, so callers get a clear signal instead of a hard\n * `400` from the gateway.\n */\nexport function applyNeonCapabilities(\n\tmodelId: string,\n\toptions: LanguageModelV3CallOptions,\n): { options: LanguageModelV3CallOptions; warnings: SharedV3Warning[] } {\n\tconst caps = getNeonModelCapabilities(modelId);\n\tconst warnings: SharedV3Warning[] = [];\n\tconst patch: Partial<LanguageModelV3CallOptions> = {};\n\n\t// The AI SDK prints the provider, the model id and the feature name ahead of\n\t// `details`, so each reason says only what those cannot: that the call still\n\t// went out, and what to reach for instead. Restating the feature name or\n\t// asserting a 400 for the unrecognised case is the overstatement this\n\t// wording exists to avoid.\n\tconst drop = (\n\t\tfeature: string,\n\t\tdetails: string,\n\t\ttype: \"unsupported\" | \"compatibility\" = \"unsupported\",\n\t) => {\n\t\twarnings.push({ type, feature, details });\n\t};\n\n\t// Precautionary rather than measured, so it is flagged as a compatibility\n\t// choice: the SDK renders `unsupported` as a flat \"is not supported\", which\n\t// would contradict the hedge in the sentence that follows it.\n\tconst unrecognizedClaude = caps.claudeSamplingUnrecognized === true;\n\tconst samplingDetails = unrecognizedClaude\n\t\t? \"This Claude version was not recognised. Claude 4.7 and newer reject sampling parameters, so the request was sent without it as a precaution; set `providerOptions.anthropic.effort` instead.\"\n\t\t: caps.family === \"anthropic\"\n\t\t\t? \"The request was sent without it. Claude 4.7 and newer reject sampling parameters; use `providerOptions.anthropic.effort` (low | medium | high | xhigh | max) to steer these models.\"\n\t\t\t: GENERIC_DETAILS;\n\tconst samplingType = unrecognizedClaude ? \"compatibility\" : \"unsupported\";\n\n\tif (options.temperature != null && !caps.supportsTemperature) {\n\t\tpatch.temperature = undefined;\n\t\tdrop(\"temperature\", samplingDetails, samplingType);\n\t}\n\tif (options.topP != null && !caps.supportsTopP) {\n\t\tpatch.topP = undefined;\n\t\tdrop(\"topP\", samplingDetails, samplingType);\n\t}\n\n\t// Anthropic-style models accept only one of temperature / topP.\n\tconst effectiveTemperature =\n\t\t\"temperature\" in patch ? patch.temperature : options.temperature;\n\tconst effectiveTopP = \"topP\" in patch ? patch.topP : options.topP;\n\tif (\n\t\tcaps.temperatureTopPMutuallyExclusive &&\n\t\teffectiveTemperature != null &&\n\t\teffectiveTopP != null\n\t) {\n\t\tpatch.topP = undefined;\n\t\twarnings.push({\n\t\t\ttype: \"compatibility\",\n\t\t\tfeature: \"topP\",\n\t\t\tdetails:\n\t\t\t\t\"Anthropic models accept only one of temperature or topP; topP was dropped and temperature kept.\",\n\t\t});\n\t}\n\n\tif (options.frequencyPenalty != null && !caps.supportsPenalties) {\n\t\tpatch.frequencyPenalty = undefined;\n\t\tdrop(\"frequencyPenalty\", PENALTY_DETAILS);\n\t}\n\tif (options.presencePenalty != null && !caps.supportsPenalties) {\n\t\tpatch.presencePenalty = undefined;\n\t\tdrop(\"presencePenalty\", PENALTY_DETAILS);\n\t}\n\tif (options.seed != null && !caps.supportsSeed) {\n\t\tpatch.seed = undefined;\n\t\tdrop(\"seed\", GENERIC_DETAILS);\n\t}\n\n\tif (options.stopSequences != null && !caps.supportsStopSequences) {\n\t\tpatch.stopSequences = undefined;\n\t\tdrop(\"stopSequences\", GENERIC_DETAILS);\n\t}\n\n\tif (!caps.supportsReasoningEffort) {\n\t\tconst { providerOptions } = options;\n\t\tconst hasProviderEffort =\n\t\t\tproviderOptions != null &&\n\t\t\tObject.values(providerOptions).some(\n\t\t\t\t(group) =>\n\t\t\t\t\t\"reasoningEffort\" in group && group.reasoningEffort != null,\n\t\t\t);\n\n\t\tif (hasProviderEffort) {\n\t\t\tdrop(\"reasoningEffort\", REASONING_EFFORT_DETAILS);\n\t\t\tif (providerOptions != null) {\n\t\t\t\tconst cleaned: SharedV3ProviderOptions = {};\n\t\t\t\tfor (const [key, group] of Object.entries(providerOptions)) {\n\t\t\t\t\tif (\"reasoningEffort\" in group) {\n\t\t\t\t\t\tconst { reasoningEffort: _removed, ...rest } = group;\n\t\t\t\t\t\tcleaned[key] = rest;\n\t\t\t\t\t} else {\n\t\t\t\t\t\tcleaned[key] = group;\n\t\t\t\t\t}\n\t\t\t\t}\n\t\t\t\tpatch.providerOptions = cleaned;\n\t\t\t}\n\t\t}\n\t}\n\n\tif (Object.keys(patch).length === 0) {\n\t\treturn { options, warnings };\n\t}\n\treturn { options: { ...options, ...patch }, warnings };\n}\n\n/**\n * Merge additional warnings into the `stream-start` part of a model stream.\n */\nexport function mergeStreamStartWarnings(extra: SharedV3Warning[]) {\n\tlet merged = false;\n\treturn new TransformStream<\n\t\tLanguageModelV3StreamPart,\n\t\tLanguageModelV3StreamPart\n\t>({\n\t\ttransform(part, controller) {\n\t\t\tif (!merged && part.type === \"stream-start\") {\n\t\t\t\tmerged = true;\n\t\t\t\tcontroller.enqueue({\n\t\t\t\t\t...part,\n\t\t\t\t\twarnings: [...extra, ...part.warnings],\n\t\t\t\t});\n\t\t\t} else {\n\t\t\t\tcontroller.enqueue(part);\n\t\t\t}\n\t\t},\n\t});\n}\n"],"mappings":";;AAQA,MAAM,kBACL;AACD,MAAM,kBACL;AACD,MAAM,2BACL;;;;;;AAOD,SAAgB,sBACf,SACA,SACuE;CACvE,MAAM,OAAO,yBAAyB,OAAO;CAC7C,MAAM,WAA8B,CAAC;CACrC,MAAM,QAA6C,CAAC;CAOpD,MAAM,QACL,SACA,SACA,OAAwC,kBACpC;EACJ,SAAS,KAAK;GAAE;GAAM;GAAS;EAAQ,CAAC;CACzC;CAKA,MAAM,qBAAqB,KAAK,+BAA+B;CAC/D,MAAM,kBAAkB,qBACrB,iMACA,KAAK,WAAW,cACf,wLACA;CACJ,MAAM,eAAe,qBAAqB,kBAAkB;CAE5D,IAAI,QAAQ,eAAe,QAAQ,CAAC,KAAK,qBAAqB;EAC7D,MAAM,cAAc,KAAA;EACpB,KAAK,eAAe,iBAAiB,YAAY;CAClD;CACA,IAAI,QAAQ,QAAQ,QAAQ,CAAC,KAAK,cAAc;EAC/C,MAAM,OAAO,KAAA;EACb,KAAK,QAAQ,iBAAiB,YAAY;CAC3C;CAGA,MAAM,uBACL,iBAAiB,QAAQ,MAAM,cAAc,QAAQ;CACtD,MAAM,gBAAgB,UAAU,QAAQ,MAAM,OAAO,QAAQ;CAC7D,IACC,KAAK,oCACL,wBAAwB,QACxB,iBAAiB,MAChB;EACD,MAAM,OAAO,KAAA;EACb,SAAS,KAAK;GACb,MAAM;GACN,SAAS;GACT,SACC;EACF,CAAC;CACF;CAEA,IAAI,QAAQ,oBAAoB,QAAQ,CAAC,KAAK,mBAAmB;EAChE,MAAM,mBAAmB,KAAA;EACzB,KAAK,oBAAoB,eAAe;CACzC;CACA,IAAI,QAAQ,mBAAmB,QAAQ,CAAC,KAAK,mBAAmB;EAC/D,MAAM,kBAAkB,KAAA;EACxB,KAAK,mBAAmB,eAAe;CACxC;CACA,IAAI,QAAQ,QAAQ,QAAQ,CAAC,KAAK,cAAc;EAC/C,MAAM,OAAO,KAAA;EACb,KAAK,QAAQ,eAAe;CAC7B;CAEA,IAAI,QAAQ,iBAAiB,QAAQ,CAAC,KAAK,uBAAuB;EACjE,MAAM,gBAAgB,KAAA;EACtB,KAAK,iBAAiB,eAAe;CACtC;CAEA,IAAI,CAAC,KAAK,yBAAyB;EAClC,MAAM,EAAE,oBAAoB;EAQ5B,IANC,mBAAmB,QACnB,OAAO,OAAO,eAAe,CAAC,CAAC,MAC7B,UACA,qBAAqB,SAAS,MAAM,mBAAmB,IACzD,GAEsB;GACtB,KAAK,mBAAmB,wBAAwB;GAChD,IAAI,mBAAmB,MAAM;IAC5B,MAAM,UAAmC,CAAC;IAC1C,KAAK,MAAM,CAAC,KAAK,UAAU,OAAO,QAAQ,eAAe,GACxD,IAAI,qBAAqB,OAAO;KAC/B,MAAM,EAAE,iBAAiB,UAAU,GAAG,SAAS;KAC/C,QAAQ,OAAO;IAChB,OACC,QAAQ,OAAO;IAGjB,MAAM,kBAAkB;GACzB;EACD;CACD;CAEA,IAAI,OAAO,KAAK,KAAK,CAAC,CAAC,WAAW,GACjC,OAAO;EAAE;EAAS;CAAS;CAE5B,OAAO;EAAE,SAAS;GAAE,GAAG;GAAS,GAAG;EAAM;EAAG;CAAS;AACtD;;;;AAKA,SAAgB,yBAAyB,OAA0B;CAClE,IAAI,SAAS;CACb,OAAO,IAAI,gBAGT,EACD,UAAU,MAAM,YAAY;EAC3B,IAAI,CAAC,UAAU,KAAK,SAAS,gBAAgB;GAC5C,SAAS;GACT,WAAW,QAAQ;IAClB,GAAG;IACH,UAAU,CAAC,GAAG,OAAO,GAAG,KAAK,QAAQ;GACtC,CAAC;EACF,OACC,WAAW,QAAQ,IAAI;CAEzB,EACD,CAAC;AACF"}
|
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
//#region src/lib/neon-chat-options.d.ts
|
|
2
|
-
/**
|
|
3
|
-
declare const NEON_MODELS_DEV_IDS: readonly ["claude-fable-5", "claude-haiku-4-5", "claude-opus-4-1", "claude-opus-4-5", "claude-opus-4-6", "claude-opus-4-7", "claude-opus-4-8", "claude-opus-5", "claude-sonnet-4-5", "claude-sonnet-4-6", "claude-sonnet-5", "gpt-5", "gpt-5-1", "gpt-5-2", "gpt-5-3-codex", "gpt-5-4", "gpt-5-4-mini", "gpt-5-4-nano", "gpt-5-5", "gpt-5-5-pro", "gpt-5-6-luna", "gpt-5-6-sol", "gpt-5-6-terra", "gpt-5-mini", "gpt-5-nano", "gpt-oss-120b", "gpt-oss-20b", "gemini-3-1-flash-lite", "gemini-3-1-pro", "gemini-3-5-flash", "gemini-3-flash", "gemma-3-12b", "llama-4-maverick", "meta-llama-3-1-8b-instruct", "meta-llama-3-3-70b-instruct", "qwen3-next-80b-a3b-instruct", "qwen35-122b-a10b", "glm-5-2", "inkling"];
|
|
2
|
+
/** Published catalog ids from https://neon.com/models.json (canonical, unprefixed). */
|
|
3
|
+
declare const NEON_MODELS_DEV_IDS: readonly ["claude-fable-5", "claude-fable-5-1", "claude-haiku-4-5", "claude-opus-4-1", "claude-opus-4-5", "claude-opus-4-6", "claude-opus-4-7", "claude-opus-4-8", "claude-opus-5", "claude-sonnet-4-5", "claude-sonnet-4-6", "claude-sonnet-5", "gpt-5", "gpt-5-1", "gpt-5-2", "gpt-5-3-codex", "gpt-5-4", "gpt-5-4-mini", "gpt-5-4-nano", "gpt-5-5", "gpt-5-5-pro", "gpt-5-6-luna", "gpt-5-6-sol", "gpt-5-6-terra", "gpt-5-mini", "gpt-5-nano", "gpt-6-astra", "gpt-oss-120b", "gpt-oss-20b", "gemini-3-1-flash-lite", "gemini-3-1-pro", "gemini-3-5-flash", "gemini-3-5-flash-lite", "gemini-3-6-flash", "gemini-3-flash", "gemma-3-12b", "llama-4-maverick", "meta-llama-3-1-8b-instruct", "meta-llama-3-3-70b-instruct", "qwen3-next-80b-a3b-instruct", "qwen35-122b-a10b", "glm-5-2", "glm-5-3-flash", "grok-4-6", "inkling", "kimi-k3"];
|
|
4
4
|
/** A known Neon AI Gateway model id. */
|
|
5
5
|
type NeonKnownModelId = (typeof NEON_MODELS_DEV_IDS)[number];
|
|
6
6
|
/**
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"neon-chat-options.d.ts","names":[],"sources":["../../src/lib/neon-chat-options.ts"],"mappings":";;
|
|
1
|
+
{"version":3,"file":"neon-chat-options.d.ts","names":[],"sources":["../../src/lib/neon-chat-options.ts"],"mappings":";;AAgBa,cAAA,mBAyDH,EAAA,SAAA,CAAA,gBAAA,EAAA,kBAAA,EAAA,kBAAA,EAAA,iBAAA,EAAA,iBAAA,EAAA,iBAAA,EAAA,iBAAA,EAAA,iBAAA,EAAA,eAAA,EAAA,mBAAA,EAAA,mBAAA,EAAA,iBAAA,EAAA,OAAA,EAAA,SAAA,EAAA,SAAA,EAAA,eAAA,EAAA,SAAA,EAAA,cAAA,EAAA,cAAA,EAAA,SAAA,EAAA,aAAA,EAAA,cAAA,EAAA,aAAA,EAAA,eAAA,EAAA,YAAA,EAAA,YAAA,EAAA,aAAA,EAAA,cAAA,EAAA,aAAA,EAAA,uBAAA,EAAA,gBAAA,EAAA,kBAAA,EAAA,uBAAA,EAAA,kBAAA,EAAA,gBAAA,EAAA,aAAA,EAAA,kBAAA,EAAA,4BAAA,EAAA,6BAAA,EAAA,6BAAA,EAAA,kBAAA,EAAA,SAAA,EAAA,eAAA,EAAA,UAAA,EAAA,SAAA,EAAA,SAAA,CAAA;AAGV;AAOY,KAPA,gBAAA,GAOe,CAAG,OAPS,mBAOO,CAAA,CAAA,MAAA,CAAA;;;;;;KAAlC,eAAA,GAAkB"}
|
|
@@ -1,7 +1,8 @@
|
|
|
1
1
|
//#region src/lib/neon-chat-options.ts
|
|
2
|
-
/**
|
|
2
|
+
/** Published catalog ids from https://neon.com/models.json (canonical, unprefixed). */
|
|
3
3
|
const NEON_MODELS_DEV_IDS = [
|
|
4
4
|
"claude-fable-5",
|
|
5
|
+
"claude-fable-5-1",
|
|
5
6
|
"claude-haiku-4-5",
|
|
6
7
|
"claude-opus-4-1",
|
|
7
8
|
"claude-opus-4-5",
|
|
@@ -26,11 +27,14 @@ const NEON_MODELS_DEV_IDS = [
|
|
|
26
27
|
"gpt-5-6-terra",
|
|
27
28
|
"gpt-5-mini",
|
|
28
29
|
"gpt-5-nano",
|
|
30
|
+
"gpt-6-astra",
|
|
29
31
|
"gpt-oss-120b",
|
|
30
32
|
"gpt-oss-20b",
|
|
31
33
|
"gemini-3-1-flash-lite",
|
|
32
34
|
"gemini-3-1-pro",
|
|
33
35
|
"gemini-3-5-flash",
|
|
36
|
+
"gemini-3-5-flash-lite",
|
|
37
|
+
"gemini-3-6-flash",
|
|
34
38
|
"gemini-3-flash",
|
|
35
39
|
"gemma-3-12b",
|
|
36
40
|
"llama-4-maverick",
|
|
@@ -39,7 +43,10 @@ const NEON_MODELS_DEV_IDS = [
|
|
|
39
43
|
"qwen3-next-80b-a3b-instruct",
|
|
40
44
|
"qwen35-122b-a10b",
|
|
41
45
|
"glm-5-2",
|
|
42
|
-
"
|
|
46
|
+
"glm-5-3-flash",
|
|
47
|
+
"grok-4-6",
|
|
48
|
+
"inkling",
|
|
49
|
+
"kimi-k3"
|
|
43
50
|
];
|
|
44
51
|
//#endregion
|
|
45
52
|
export { NEON_MODELS_DEV_IDS };
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"neon-chat-options.js","names":[],"sources":["../../src/lib/neon-chat-options.ts"],"sourcesContent":["// Models served by the Neon AI Gateway. Anthropic and OpenAI models are routed\n// to their provider-native endpoints; everything else uses the unified MLflow\n// (OpenAI-compatible) endpoint.\n//\n// Ids use the canonical Neon (unprefixed) form, matching
|
|
1
|
+
{"version":3,"file":"neon-chat-options.js","names":[],"sources":["../../src/lib/neon-chat-options.ts"],"sourcesContent":["// Models served by the Neon AI Gateway. Anthropic and OpenAI models are routed\n// to their provider-native endpoints; everything else uses the unified MLflow\n// (OpenAI-compatible) endpoint.\n//\n// Ids use the canonical Neon (unprefixed) form, matching\n// https://neon.com/models.json. The gateway also accepts the legacy\n// `databricks-` prefixed form, which still resolves via the `(string & {})`\n// fallback on `NeonChatModelId`.\n//\n// `NEON_MODELS_DEV_IDS` mirrors neon.com/models.json `neon.models` exactly —\n// `neon-catalog-drift.test.ts` fails if the two diverge. A branch's live list\n// is `GET $NEON_AI_GATEWAY_BASE_URL/v1/models`. There is deliberately no second\n// list for ids the gateway serves ahead of the published catalog: such an id\n// already works because `NeonChatModelId` accepts any string.\n\n/** Published catalog ids from https://neon.com/models.json (canonical, unprefixed). */\nexport const NEON_MODELS_DEV_IDS = [\n\t// Anthropic (native Messages API)\n\t\"claude-fable-5\",\n\t\"claude-fable-5-1\",\n\t\"claude-haiku-4-5\",\n\t\"claude-opus-4-1\",\n\t\"claude-opus-4-5\",\n\t\"claude-opus-4-6\",\n\t\"claude-opus-4-7\",\n\t\"claude-opus-4-8\",\n\t\"claude-opus-5\",\n\t\"claude-sonnet-4-5\",\n\t\"claude-sonnet-4-6\",\n\t\"claude-sonnet-5\",\n\t// OpenAI (native Responses API)\n\t\"gpt-5\",\n\t\"gpt-5-1\",\n\t\"gpt-5-2\",\n\t\"gpt-5-3-codex\",\n\t\"gpt-5-4\",\n\t\"gpt-5-4-mini\",\n\t\"gpt-5-4-nano\",\n\t\"gpt-5-5\",\n\t\"gpt-5-5-pro\",\n\t\"gpt-5-6-luna\",\n\t\"gpt-5-6-sol\",\n\t\"gpt-5-6-terra\",\n\t\"gpt-5-mini\",\n\t\"gpt-5-nano\",\n\t\"gpt-6-astra\",\n\t// OpenAI open-weight (unified MLflow endpoint)\n\t\"gpt-oss-120b\",\n\t\"gpt-oss-20b\",\n\t// Google (unified MLflow endpoint)\n\t\"gemini-3-1-flash-lite\",\n\t\"gemini-3-1-pro\",\n\t\"gemini-3-5-flash\",\n\t\"gemini-3-5-flash-lite\",\n\t\"gemini-3-6-flash\",\n\t\"gemini-3-flash\",\n\t\"gemma-3-12b\",\n\t// Meta (unified MLflow endpoint)\n\t\"llama-4-maverick\",\n\t\"meta-llama-3-1-8b-instruct\",\n\t\"meta-llama-3-3-70b-instruct\",\n\t// Alibaba (unified MLflow endpoint)\n\t\"qwen3-next-80b-a3b-instruct\",\n\t\"qwen35-122b-a10b\",\n\t// Zhipu (unified MLflow endpoint)\n\t\"glm-5-2\",\n\t\"glm-5-3-flash\",\n\t// xAI (unified MLflow endpoint)\n\t\"grok-4-6\",\n\t// Thinking Machines (unified MLflow endpoint)\n\t\"inkling\",\n\t// Moonshot (unified MLflow endpoint)\n\t\"kimi-k3\",\n] as const;\n\n/** A known Neon AI Gateway model id. */\nexport type NeonKnownModelId = (typeof NEON_MODELS_DEV_IDS)[number];\n\n/**\n * A Neon AI Gateway model id. Known ids are listed for autocomplete; any other\n * id (e.g. a brand-new model, or the legacy `databricks-` prefixed form) is\n * accepted via the `(string & {})` fallback.\n */\nexport type NeonChatModelId = NeonKnownModelId | (string & {});\n"],"mappings":";;AAgBA,MAAa,sBAAsB;CAElC;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CAEA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CAEA;CACA;CAEA;CACA;CACA;CACA;CACA;CACA;CACA;CAEA;CACA;CACA;CAEA;CACA;CAEA;CACA;CAEA;CAEA;CAEA;AACD"}
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"neon-model-capabilities.d.ts","names":[],"sources":["../../src/lib/neon-model-capabilities.ts"],"mappings":";KAAY,eAAA;AAAZ;AAiBA;AAEA;AAiBA;
|
|
1
|
+
{"version":3,"file":"neon-model-capabilities.d.ts","names":[],"sources":["../../src/lib/neon-model-capabilities.ts"],"mappings":";KAAY,eAAA;AAAZ;AAiBA;AAEA;AAiBA;AAiGA;;;;;;KApHY,cAAA;iBAEI,iBAAA,mBAAoC;UAiBnC,qBAAA;UACR;;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBAgGO,wBAAA,mBAEb"}
|
|
@@ -5,6 +5,34 @@ function getNeonModelRoute(modelId) {
|
|
|
5
5
|
if (id.includes("gpt-") && !id.includes("gpt-oss") || id.includes("codex")) return "openai";
|
|
6
6
|
return "mlflow";
|
|
7
7
|
}
|
|
8
|
+
/**
|
|
9
|
+
* Gemini models the gateway rejects penalties on, by measurement. Their older
|
|
10
|
+
* siblings accept them, so this cannot be derived from the id.
|
|
11
|
+
*
|
|
12
|
+
* Matched exactly rather than by substring: `gemini-3-6-flash` is a prefix of
|
|
13
|
+
* any `gemini-3-6-flash-*` id Google may ship next, and inheriting a measured
|
|
14
|
+
* restriction on the strength of a shared prefix is how the Gemini rule was
|
|
15
|
+
* wrong for these two in the first place.
|
|
16
|
+
*/
|
|
17
|
+
const GEMINI_NO_PENALTIES = /* @__PURE__ */ new Set(["gemini-3-5-flash-lite", "gemini-3-6-flash"]);
|
|
18
|
+
/** Gemini models that reject `temperature` and `topP` outright. */
|
|
19
|
+
const GEMINI_NO_SAMPLING = /* @__PURE__ */ new Set(["gemini-3-6-flash"]);
|
|
20
|
+
/**
|
|
21
|
+
* GPT-5 ids that reject `temperature` despite carrying a minor version.
|
|
22
|
+
*
|
|
23
|
+
* The version rule below reads `gpt-5-5-pro` as 5.5 and concludes it takes
|
|
24
|
+
* sampling parameters. It does not — the Responses API answers
|
|
25
|
+
* `Unsupported parameter: temperature`.
|
|
26
|
+
*/
|
|
27
|
+
const GPT5_NO_SAMPLING = /* @__PURE__ */ new Set(["gpt-5-5-pro"]);
|
|
28
|
+
/**
|
|
29
|
+
* The gateway accepts an optional `databricks-` prefix on any id, so strip it
|
|
30
|
+
* before an exact-match lookup. Prefix rules elsewhere in this file tolerate it
|
|
31
|
+
* incidentally; exact matches do not.
|
|
32
|
+
*/
|
|
33
|
+
function canonicalId(id) {
|
|
34
|
+
return id.startsWith("databricks-") ? id.slice(11) : id;
|
|
35
|
+
}
|
|
8
36
|
const PERMISSIVE = {
|
|
9
37
|
supportsTemperature: true,
|
|
10
38
|
supportsTopP: true,
|
|
@@ -54,13 +82,19 @@ function getNeonModelCapabilities(modelId) {
|
|
|
54
82
|
...version === null ? { claudeSamplingUnrecognized: true } : {}
|
|
55
83
|
};
|
|
56
84
|
}
|
|
57
|
-
if (id.includes("gemini"))
|
|
58
|
-
|
|
59
|
-
|
|
60
|
-
|
|
61
|
-
|
|
85
|
+
if (id.includes("gemini")) {
|
|
86
|
+
const canonical = canonicalId(id);
|
|
87
|
+
const acceptsSampling = !GEMINI_NO_SAMPLING.has(canonical);
|
|
88
|
+
return {
|
|
89
|
+
family: "google",
|
|
90
|
+
...PERMISSIVE,
|
|
91
|
+
supportsTemperature: acceptsSampling,
|
|
92
|
+
supportsTopP: acceptsSampling,
|
|
93
|
+
supportsPenalties: !GEMINI_NO_PENALTIES.has(canonical)
|
|
94
|
+
};
|
|
95
|
+
}
|
|
62
96
|
if (/gpt-5/.test(id)) {
|
|
63
|
-
const hasMinorVersion = /gpt-5[.-]\d/.test(id);
|
|
97
|
+
const hasMinorVersion = /gpt-5[.-]\d/.test(id) && !GPT5_NO_SAMPLING.has(canonicalId(id));
|
|
64
98
|
return {
|
|
65
99
|
family: "openai",
|
|
66
100
|
supportsTemperature: hasMinorVersion,
|
|
@@ -72,6 +106,16 @@ function getNeonModelCapabilities(modelId) {
|
|
|
72
106
|
supportsReasoningEffort: true
|
|
73
107
|
};
|
|
74
108
|
}
|
|
109
|
+
if (canonicalId(id) === "gpt-6-astra") return {
|
|
110
|
+
family: "openai",
|
|
111
|
+
supportsTemperature: false,
|
|
112
|
+
supportsTopP: true,
|
|
113
|
+
temperatureTopPMutuallyExclusive: false,
|
|
114
|
+
supportsPenalties: true,
|
|
115
|
+
supportsSeed: true,
|
|
116
|
+
supportsStopSequences: true,
|
|
117
|
+
supportsReasoningEffort: true
|
|
118
|
+
};
|
|
75
119
|
if (id.includes("llama")) return {
|
|
76
120
|
family: "meta",
|
|
77
121
|
supportsTemperature: true,
|
|
@@ -95,7 +139,7 @@ function getNeonModelCapabilities(modelId) {
|
|
|
95
139
|
supportsPenalties: false,
|
|
96
140
|
supportsSeed: false
|
|
97
141
|
};
|
|
98
|
-
if (id.includes("glm") || id.includes("inkling")) return {
|
|
142
|
+
if (id.includes("glm") || id.includes("inkling") || id.includes("kimi")) return {
|
|
99
143
|
family: "other",
|
|
100
144
|
...PERMISSIVE,
|
|
101
145
|
supportsPenalties: false
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"neon-model-capabilities.js","names":[],"sources":["../../src/lib/neon-model-capabilities.ts"],"sourcesContent":["export type NeonModelFamily =\n\t| \"anthropic\"\n\t| \"google\"\n\t| \"openai\"\n\t| \"meta\"\n\t| \"other\";\n\n/**\n * Which gateway endpoint a model should be routed to.\n *\n * - `anthropic`: native Messages API — streaming structured output + native reasoning.\n * - `openai`: native Responses API — models served only natively (e.g. Codex),\n * native reasoning, and the image-generation tool.\n * - `mlflow`: the unified, OpenAI-compatible endpoint (the fallback). Gemini is\n * routed here because the gateway's native Gemini endpoint does not support\n * streaming (`streamGenerateContent` is rejected).\n */\nexport type NeonModelRoute = \"anthropic\" | \"openai\" | \"mlflow\";\n\nexport function getNeonModelRoute(modelId: string): NeonModelRoute {\n\tconst id = modelId.toLowerCase();\n\tif (id.includes(\"claude\")) {\n\t\treturn \"anthropic\";\n\t}\n\t// OpenAI proprietary models (gpt-4/gpt-5 families, Codex) are served only via\n\t// the native Responses API. `gpt-oss` is open-weight and served on the\n\t// unified chat endpoint, so it is intentionally excluded here.\n\tif (\n\t\t(id.includes(\"gpt-\") && !id.includes(\"gpt-oss\")) ||\n\t\tid.includes(\"codex\")\n\t) {\n\t\treturn \"openai\";\n\t}\n\treturn \"mlflow\";\n}\n\nexport interface NeonModelCapabilities {\n\tfamily: NeonModelFamily;\n\tsupportsTemperature: boolean;\n\tsupportsTopP: boolean;\n\t/**\n\t * Whether `temperature` and `topP` may be sent together. Anthropic models\n\t * accept only one of the two.\n\t */\n\ttemperatureTopPMutuallyExclusive: boolean;\n\tsupportsPenalties: boolean;\n\tsupportsSeed: boolean;\n\tsupportsStopSequences: boolean;\n\tsupportsReasoningEffort: boolean;\n\t/**\n\t * True when a `claude-*` id could not be parsed into a version and was\n\t * assumed to be new. The sampling restriction is then a precaution rather\n\t * than something measured, and the warning says so instead of asserting a\n\t * capability nobody checked.\n\t */\n\tclaudeSamplingUnrecognized?: boolean;\n}\n\nconst PERMISSIVE: Omit<NeonModelCapabilities, \"family\"> = {\n\tsupportsTemperature: true,\n\tsupportsTopP: true,\n\ttemperatureTopPMutuallyExclusive: false,\n\tsupportsPenalties: true,\n\tsupportsSeed: true,\n\tsupportsStopSequences: true,\n\tsupportsReasoningEffort: true,\n};\n\n/**\n * Whether a Claude id predates the 4.7 cutoff where the gateway stopped\n * accepting `temperature` and `top_p`. Ids look like `claude-<tier>-<major>` or\n * `claude-<tier>-<major>-<minor>`, optionally behind the legacy `databricks-`\n * prefix. An id we cannot parse is treated as new, which is the safe direction.\n */\nfunction claudeVersion(id: string): { major: number; minor: number } | null {\n\t// Anchored: `claude-opus-4-beta` and `claude-opus-4.7` must not match their\n\t// leading digits and be read as old models. A dated id such as\n\t// `claude-3-5-sonnet-20241022` does not parse either, and is reported as\n\t// unrecognised rather than silently classified.\n\tconst match = /^(?:databricks-)?claude-[a-z]+-(\\d+)(?:-(\\d+))?$/.exec(id);\n\tif (!match) {\n\t\treturn null;\n\t}\n\treturn {\n\t\tmajor: Number(match[1]),\n\t\tminor: match[2] === undefined ? 0 : Number(match[2]),\n\t};\n}\n\n/**\n * Heuristic, prefix-based capability detection for MLflow-routed Neon models.\n *\n * The unified endpoint proxies to heterogeneous upstream providers, each of\n * which accepts a different subset of the OpenAI-style parameters the AI SDK\n * emits. Sending a parameter an upstream rejects results in a hard `400`, so we\n * strip the parameters a family is known to reject and surface a warning\n * instead. Unknown/untested models stay permissive (passed through unchanged).\n */\nexport function getNeonModelCapabilities(\n\tmodelId: string,\n): NeonModelCapabilities {\n\tconst id = modelId.toLowerCase();\n\n\t// Anthropic (Claude): rejects penalties and seed, accepts only one of\n\t// temperature/topP, and rejects the OpenAI `reasoning_effort` field.\n\t//\n\t// From Claude 4.7 onward the gateway also rejects non-default sampling\n\t// entirely — `does not support the temperature parameter` /\n\t// `does not support sampling parameters: top_p` — because those models are\n\t// steered with `output_config.effort` instead. Measured: 4.1/4.5/4.6 accept\n\t// both, 4.7/4.8/5 reject both, and every id accepts the default\n\t// `temperature: 1.0`. A version comparison rather than a list of ids, so a\n\t// future Claude inherits the restriction: dropping a parameter the model\n\t// would have accepted costs a warning, claiming one it rejects costs a 400.\n\tif (id.includes(\"claude\")) {\n\t\tconst version = claudeVersion(id);\n\t\tconst sampling =\n\t\t\tversion !== null &&\n\t\t\t(version.major < 4 || (version.major === 4 && version.minor <= 6));\n\t\treturn {\n\t\t\tfamily: \"anthropic\",\n\t\t\tsupportsTemperature: sampling,\n\t\t\tsupportsTopP: sampling,\n\t\t\ttemperatureTopPMutuallyExclusive: true,\n\t\t\tsupportsPenalties: false,\n\t\t\tsupportsSeed: false,\n\t\t\tsupportsStopSequences: true,\n\t\t\tsupportsReasoningEffort: false,\n\t\t\t...(version === null ? { claudeSamplingUnrecognized: true } : {}),\n\t\t};\n\t}\n\n\t// Google (Gemini): accepts standard sampling params but rejects the OpenAI\n\t// `reasoning_effort` field (not part of Gemini's generation config).\n\tif (id.includes(\"gemini\")) {\n\t\treturn {\n\t\t\tfamily: \"google\",\n\t\t\t...PERMISSIVE,\n\t\t\tsupportsReasoningEffort: false,\n\t\t};\n\t}\n\n\t// OpenAI GPT-5 reasoning family (routed to the native Responses API). The\n\t// Responses model strips parameters the Responses API doesn't accept\n\t// (penalties, seed, stop), but its reasoning-model detection matches the bare\n\t// model id (`gpt-5`), which the gateway's optional `databricks-` prefix can\n\t// defeat. So we only handle the temperature/topP restriction here: the\n\t// original gpt-5 / gpt-5-mini / gpt-5-nano require the default temperature and\n\t// reject topP, while gpt-5.1+ (a minor version digit follows) accept them\n\t// again. The regex matches both prefixed and unprefixed ids.\n\tif (/gpt-5/.test(id)) {\n\t\tconst hasMinorVersion = /gpt-5[.-]\\d/.test(id);\n\t\treturn {\n\t\t\tfamily: \"openai\",\n\t\t\tsupportsTemperature: hasMinorVersion,\n\t\t\tsupportsTopP: hasMinorVersion,\n\t\t\ttemperatureTopPMutuallyExclusive: false,\n\t\t\tsupportsPenalties: true,\n\t\t\tsupportsSeed: true,\n\t\t\tsupportsStopSequences: true,\n\t\t\tsupportsReasoningEffort: true,\n\t\t};\n\t}\n\n\t// Meta (Llama): rejects penalties and seed; accepts sampling params and stop.\n\tif (id.includes(\"llama\")) {\n\t\treturn {\n\t\t\tfamily: \"meta\",\n\t\t\tsupportsTemperature: true,\n\t\t\tsupportsTopP: true,\n\t\t\ttemperatureTopPMutuallyExclusive: false,\n\t\t\tsupportsPenalties: false,\n\t\t\tsupportsSeed: false,\n\t\t\tsupportsStopSequences: true,\n\t\t\tsupportsReasoningEffort: true,\n\t\t};\n\t}\n\n\t// Everything else on the unified endpoint rejects penalties with\n\t// `parameter \"frequency_penalty\" must be equal to 0`. Gemini, handled above,\n\t// is the only MLflow-routed family that accepts them. Seed and stop vary, so\n\t// each family carries what was measured rather than inheriting a guess.\n\tif (id.includes(\"gpt-oss\")) {\n\t\treturn {\n\t\t\tfamily: \"other\",\n\t\t\t...PERMISSIVE,\n\t\t\tsupportsPenalties: false,\n\t\t\tsupportsSeed: false,\n\t\t\tsupportsStopSequences: false,\n\t\t};\n\t}\n\n\tif (id.includes(\"qwen\") || id.includes(\"gemma\")) {\n\t\treturn {\n\t\t\tfamily: \"other\",\n\t\t\t...PERMISSIVE,\n\t\t\tsupportsPenalties: false,\n\t\t\tsupportsSeed: false,\n\t\t};\n\t}\n\n\tif (id.includes(\"glm\") || id.includes(\"inkling\")) {\n\t\treturn { family: \"other\", ...PERMISSIVE, supportsPenalties: false };\n\t}\n\n\treturn { family: \"other\", ...PERMISSIVE };\n}\n"],"mappings":";AAmBA,SAAgB,kBAAkB,SAAiC;CAClE,MAAM,KAAK,QAAQ,YAAY;CAC/B,IAAI,GAAG,SAAS,QAAQ,GACvB,OAAO;CAKR,IACE,GAAG,SAAS,MAAM,KAAK,CAAC,GAAG,SAAS,SAAS,KAC9C,GAAG,SAAS,OAAO,GAEnB,OAAO;CAER,OAAO;AACR;AAwBA,MAAM,aAAoD;CACzD,qBAAqB;CACrB,cAAc;CACd,kCAAkC;CAClC,mBAAmB;CACnB,cAAc;CACd,uBAAuB;CACvB,yBAAyB;AAC1B;;;;;;;AAQA,SAAS,cAAc,IAAqD;CAK3E,MAAM,QAAQ,mDAAmD,KAAK,EAAE;CACxE,IAAI,CAAC,OACJ,OAAO;CAER,OAAO;EACN,OAAO,OAAO,MAAM,EAAE;EACtB,OAAO,MAAM,OAAO,KAAA,IAAY,IAAI,OAAO,MAAM,EAAE;CACpD;AACD;;;;;;;;;;AAWA,SAAgB,yBACf,SACwB;CACxB,MAAM,KAAK,QAAQ,YAAY;CAa/B,IAAI,GAAG,SAAS,QAAQ,GAAG;EAC1B,MAAM,UAAU,cAAc,EAAE;EAChC,MAAM,WACL,YAAY,SACX,QAAQ,QAAQ,KAAM,QAAQ,UAAU,KAAK,QAAQ,SAAS;EAChE,OAAO;GACN,QAAQ;GACR,qBAAqB;GACrB,cAAc;GACd,kCAAkC;GAClC,mBAAmB;GACnB,cAAc;GACd,uBAAuB;GACvB,yBAAyB;GACzB,GAAI,YAAY,OAAO,EAAE,4BAA4B,KAAK,IAAI,CAAC;EAChE;CACD;CAIA,IAAI,GAAG,SAAS,QAAQ,GACvB,OAAO;EACN,QAAQ;EACR,GAAG;EACH,yBAAyB;CAC1B;CAWD,IAAI,QAAQ,KAAK,EAAE,GAAG;EACrB,MAAM,kBAAkB,cAAc,KAAK,EAAE;EAC7C,OAAO;GACN,QAAQ;GACR,qBAAqB;GACrB,cAAc;GACd,kCAAkC;GAClC,mBAAmB;GACnB,cAAc;GACd,uBAAuB;GACvB,yBAAyB;EAC1B;CACD;CAGA,IAAI,GAAG,SAAS,OAAO,GACtB,OAAO;EACN,QAAQ;EACR,qBAAqB;EACrB,cAAc;EACd,kCAAkC;EAClC,mBAAmB;EACnB,cAAc;EACd,uBAAuB;EACvB,yBAAyB;CAC1B;CAOD,IAAI,GAAG,SAAS,SAAS,GACxB,OAAO;EACN,QAAQ;EACR,GAAG;EACH,mBAAmB;EACnB,cAAc;EACd,uBAAuB;CACxB;CAGD,IAAI,GAAG,SAAS,MAAM,KAAK,GAAG,SAAS,OAAO,GAC7C,OAAO;EACN,QAAQ;EACR,GAAG;EACH,mBAAmB;EACnB,cAAc;CACf;CAGD,IAAI,GAAG,SAAS,KAAK,KAAK,GAAG,SAAS,SAAS,GAC9C,OAAO;EAAE,QAAQ;EAAS,GAAG;EAAY,mBAAmB;CAAM;CAGnE,OAAO;EAAE,QAAQ;EAAS,GAAG;CAAW;AACzC"}
|
|
1
|
+
{"version":3,"file":"neon-model-capabilities.js","names":[],"sources":["../../src/lib/neon-model-capabilities.ts"],"sourcesContent":["export type NeonModelFamily =\n\t| \"anthropic\"\n\t| \"google\"\n\t| \"openai\"\n\t| \"meta\"\n\t| \"other\";\n\n/**\n * Which gateway endpoint a model should be routed to.\n *\n * - `anthropic`: native Messages API — streaming structured output + native reasoning.\n * - `openai`: native Responses API — models served only natively (e.g. Codex),\n * native reasoning, and the image-generation tool.\n * - `mlflow`: the unified, OpenAI-compatible endpoint (the fallback). Gemini is\n * routed here because the gateway's native Gemini endpoint does not support\n * streaming (`streamGenerateContent` is rejected).\n */\nexport type NeonModelRoute = \"anthropic\" | \"openai\" | \"mlflow\";\n\nexport function getNeonModelRoute(modelId: string): NeonModelRoute {\n\tconst id = modelId.toLowerCase();\n\tif (id.includes(\"claude\")) {\n\t\treturn \"anthropic\";\n\t}\n\t// OpenAI proprietary models (gpt-4/gpt-5 families, Codex) are served only via\n\t// the native Responses API. `gpt-oss` is open-weight and served on the\n\t// unified chat endpoint, so it is intentionally excluded here.\n\tif (\n\t\t(id.includes(\"gpt-\") && !id.includes(\"gpt-oss\")) ||\n\t\tid.includes(\"codex\")\n\t) {\n\t\treturn \"openai\";\n\t}\n\treturn \"mlflow\";\n}\n\nexport interface NeonModelCapabilities {\n\tfamily: NeonModelFamily;\n\tsupportsTemperature: boolean;\n\tsupportsTopP: boolean;\n\t/**\n\t * Whether `temperature` and `topP` may be sent together. Anthropic models\n\t * accept only one of the two.\n\t */\n\ttemperatureTopPMutuallyExclusive: boolean;\n\tsupportsPenalties: boolean;\n\tsupportsSeed: boolean;\n\tsupportsStopSequences: boolean;\n\tsupportsReasoningEffort: boolean;\n\t/**\n\t * True when a `claude-*` id could not be parsed into a version and was\n\t * assumed to be new. The sampling restriction is then a precaution rather\n\t * than something measured, and the warning says so instead of asserting a\n\t * capability nobody checked.\n\t */\n\tclaudeSamplingUnrecognized?: boolean;\n}\n\n/**\n * Gemini models the gateway rejects penalties on, by measurement. Their older\n * siblings accept them, so this cannot be derived from the id.\n *\n * Matched exactly rather than by substring: `gemini-3-6-flash` is a prefix of\n * any `gemini-3-6-flash-*` id Google may ship next, and inheriting a measured\n * restriction on the strength of a shared prefix is how the Gemini rule was\n * wrong for these two in the first place.\n */\nconst GEMINI_NO_PENALTIES = new Set([\n\t\"gemini-3-5-flash-lite\",\n\t\"gemini-3-6-flash\",\n]);\n\n/** Gemini models that reject `temperature` and `topP` outright. */\nconst GEMINI_NO_SAMPLING = new Set([\"gemini-3-6-flash\"]);\n\n/**\n * GPT-5 ids that reject `temperature` despite carrying a minor version.\n *\n * The version rule below reads `gpt-5-5-pro` as 5.5 and concludes it takes\n * sampling parameters. It does not — the Responses API answers\n * `Unsupported parameter: temperature`.\n */\nconst GPT5_NO_SAMPLING = new Set([\"gpt-5-5-pro\"]);\n\n/**\n * The gateway accepts an optional `databricks-` prefix on any id, so strip it\n * before an exact-match lookup. Prefix rules elsewhere in this file tolerate it\n * incidentally; exact matches do not.\n */\nfunction canonicalId(id: string): string {\n\treturn id.startsWith(\"databricks-\") ? id.slice(\"databricks-\".length) : id;\n}\n\nconst PERMISSIVE: Omit<NeonModelCapabilities, \"family\"> = {\n\tsupportsTemperature: true,\n\tsupportsTopP: true,\n\ttemperatureTopPMutuallyExclusive: false,\n\tsupportsPenalties: true,\n\tsupportsSeed: true,\n\tsupportsStopSequences: true,\n\tsupportsReasoningEffort: true,\n};\n\n/**\n * Whether a Claude id predates the 4.7 cutoff where the gateway stopped\n * accepting `temperature` and `top_p`. Ids look like `claude-<tier>-<major>` or\n * `claude-<tier>-<major>-<minor>`, optionally behind the legacy `databricks-`\n * prefix. An id we cannot parse is treated as new, which is the safe direction.\n */\nfunction claudeVersion(id: string): { major: number; minor: number } | null {\n\t// Anchored: `claude-opus-4-beta` and `claude-opus-4.7` must not match their\n\t// leading digits and be read as old models. A dated id such as\n\t// `claude-3-5-sonnet-20241022` does not parse either, and is reported as\n\t// unrecognised rather than silently classified.\n\tconst match = /^(?:databricks-)?claude-[a-z]+-(\\d+)(?:-(\\d+))?$/.exec(id);\n\tif (!match) {\n\t\treturn null;\n\t}\n\treturn {\n\t\tmajor: Number(match[1]),\n\t\tminor: match[2] === undefined ? 0 : Number(match[2]),\n\t};\n}\n\n/**\n * Heuristic, prefix-based capability detection for MLflow-routed Neon models.\n *\n * The unified endpoint proxies to heterogeneous upstream providers, each of\n * which accepts a different subset of the OpenAI-style parameters the AI SDK\n * emits. Sending a parameter an upstream rejects results in a hard `400`, so we\n * strip the parameters a family is known to reject and surface a warning\n * instead. Unknown/untested models stay permissive (passed through unchanged).\n */\nexport function getNeonModelCapabilities(\n\tmodelId: string,\n): NeonModelCapabilities {\n\tconst id = modelId.toLowerCase();\n\n\t// Anthropic (Claude): rejects penalties and seed, accepts only one of\n\t// temperature/topP, and rejects the OpenAI `reasoning_effort` field.\n\t//\n\t// From Claude 4.7 onward the gateway also rejects non-default sampling\n\t// entirely — `does not support the temperature parameter` /\n\t// `does not support sampling parameters: top_p` — because those models are\n\t// steered with `output_config.effort` instead. Measured: 4.1/4.5/4.6 accept\n\t// both, 4.7/4.8/5 reject both, and every id accepts the default\n\t// `temperature: 1.0`. A version comparison rather than a list of ids, so a\n\t// future Claude inherits the restriction: dropping a parameter the model\n\t// would have accepted costs a warning, claiming one it rejects costs a 400.\n\tif (id.includes(\"claude\")) {\n\t\tconst version = claudeVersion(id);\n\t\tconst sampling =\n\t\t\tversion !== null &&\n\t\t\t(version.major < 4 || (version.major === 4 && version.minor <= 6));\n\t\treturn {\n\t\t\tfamily: \"anthropic\",\n\t\t\tsupportsTemperature: sampling,\n\t\t\tsupportsTopP: sampling,\n\t\t\ttemperatureTopPMutuallyExclusive: true,\n\t\t\tsupportsPenalties: false,\n\t\t\tsupportsSeed: false,\n\t\t\tsupportsStopSequences: true,\n\t\t\tsupportsReasoningEffort: false,\n\t\t\t...(version === null ? { claudeSamplingUnrecognized: true } : {}),\n\t\t};\n\t}\n\n\t// Google (Gemini): accepts standard sampling params, and does take\n\t// `reasoning_effort` — the gateway maps it onto Gemini's thinking config.\n\t// Measured on gemini-3-6-flash: `minimal` produces 0 reasoning tokens and\n\t// `high` produces 310, against 130 with the field absent.\n\t//\n\t// Newer Gemini models are stricter, and not in a way the id predicts: the\n\t// gateway rejects penalties on `gemini-3-5-flash-lite` while accepting them\n\t// on `gemini-3-1-flash-lite`, so neither the version nor the `-lite` suffix\n\t// is the rule. `gemini-3-6-flash` refuses temperature and topP outright\n\t// (\"does not support the temperature parameter\"). Both are listed by id\n\t// because that is what was measured; a new Gemini model inherits the\n\t// permissive default until someone measures it.\n\tif (id.includes(\"gemini\")) {\n\t\tconst canonical = canonicalId(id);\n\t\tconst acceptsSampling = !GEMINI_NO_SAMPLING.has(canonical);\n\t\treturn {\n\t\t\tfamily: \"google\",\n\t\t\t...PERMISSIVE,\n\t\t\tsupportsTemperature: acceptsSampling,\n\t\t\tsupportsTopP: acceptsSampling,\n\t\t\tsupportsPenalties: !GEMINI_NO_PENALTIES.has(canonical),\n\t\t};\n\t}\n\n\t// OpenAI GPT-5 reasoning family (routed to the native Responses API). The\n\t// Responses model strips parameters the Responses API doesn't accept\n\t// (penalties, seed, stop), but its reasoning-model detection matches the bare\n\t// model id (`gpt-5`), which the gateway's optional `databricks-` prefix can\n\t// defeat. So we only handle the temperature/topP restriction here: the\n\t// original gpt-5 / gpt-5-mini / gpt-5-nano require the default temperature and\n\t// reject topP, while gpt-5.1+ (a minor version digit follows) accept them\n\t// again. The regex matches both prefixed and unprefixed ids.\n\tif (/gpt-5/.test(id)) {\n\t\tconst hasMinorVersion =\n\t\t\t/gpt-5[.-]\\d/.test(id) && !GPT5_NO_SAMPLING.has(canonicalId(id));\n\t\treturn {\n\t\t\tfamily: \"openai\",\n\t\t\tsupportsTemperature: hasMinorVersion,\n\t\t\tsupportsTopP: hasMinorVersion,\n\t\t\ttemperatureTopPMutuallyExclusive: false,\n\t\t\tsupportsPenalties: true,\n\t\t\tsupportsSeed: true,\n\t\t\tsupportsStopSequences: true,\n\t\t\tsupportsReasoningEffort: true,\n\t\t};\n\t}\n\n\t// Catalog `temperature: false` on gpt-6-astra; the gpt-5 minor-version rule\n\t// does not match gpt-6. Exact id: a future gpt-6-* is unmeasured.\n\tif (canonicalId(id) === \"gpt-6-astra\") {\n\t\treturn {\n\t\t\tfamily: \"openai\",\n\t\t\tsupportsTemperature: false,\n\t\t\tsupportsTopP: true,\n\t\t\ttemperatureTopPMutuallyExclusive: false,\n\t\t\tsupportsPenalties: true,\n\t\t\tsupportsSeed: true,\n\t\t\tsupportsStopSequences: true,\n\t\t\tsupportsReasoningEffort: true,\n\t\t};\n\t}\n\n\t// Meta (Llama): rejects penalties and seed; accepts sampling params and stop.\n\tif (id.includes(\"llama\")) {\n\t\treturn {\n\t\t\tfamily: \"meta\",\n\t\t\tsupportsTemperature: true,\n\t\t\tsupportsTopP: true,\n\t\t\ttemperatureTopPMutuallyExclusive: false,\n\t\t\tsupportsPenalties: false,\n\t\t\tsupportsSeed: false,\n\t\t\tsupportsStopSequences: true,\n\t\t\tsupportsReasoningEffort: true,\n\t\t};\n\t}\n\n\t// Everything else on the unified endpoint rejects penalties with\n\t// `parameter \"frequency_penalty\" must be equal to 0`. Only the older Gemini\n\t// models, handled above, accept them. Seed and stop vary, so each family\n\t// carries what was measured rather than inheriting a guess.\n\tif (id.includes(\"gpt-oss\")) {\n\t\treturn {\n\t\t\tfamily: \"other\",\n\t\t\t...PERMISSIVE,\n\t\t\tsupportsPenalties: false,\n\t\t\tsupportsSeed: false,\n\t\t\tsupportsStopSequences: false,\n\t\t};\n\t}\n\n\tif (id.includes(\"qwen\") || id.includes(\"gemma\")) {\n\t\treturn {\n\t\t\tfamily: \"other\",\n\t\t\t...PERMISSIVE,\n\t\t\tsupportsPenalties: false,\n\t\t\tsupportsSeed: false,\n\t\t};\n\t}\n\n\tif (id.includes(\"glm\") || id.includes(\"inkling\") || id.includes(\"kimi\")) {\n\t\treturn { family: \"other\", ...PERMISSIVE, supportsPenalties: false };\n\t}\n\n\treturn { family: \"other\", ...PERMISSIVE };\n}\n"],"mappings":";AAmBA,SAAgB,kBAAkB,SAAiC;CAClE,MAAM,KAAK,QAAQ,YAAY;CAC/B,IAAI,GAAG,SAAS,QAAQ,GACvB,OAAO;CAKR,IACE,GAAG,SAAS,MAAM,KAAK,CAAC,GAAG,SAAS,SAAS,KAC9C,GAAG,SAAS,OAAO,GAEnB,OAAO;CAER,OAAO;AACR;;;;;;;;;;AAiCA,MAAM,sCAAsB,IAAI,IAAI,CACnC,yBACA,kBACD,CAAC;;AAGD,MAAM,qCAAqB,IAAI,IAAI,CAAC,kBAAkB,CAAC;;;;;;;;AASvD,MAAM,mCAAmB,IAAI,IAAI,CAAC,aAAa,CAAC;;;;;;AAOhD,SAAS,YAAY,IAAoB;CACxC,OAAO,GAAG,WAAW,aAAa,IAAI,GAAG,MAAM,EAAoB,IAAI;AACxE;AAEA,MAAM,aAAoD;CACzD,qBAAqB;CACrB,cAAc;CACd,kCAAkC;CAClC,mBAAmB;CACnB,cAAc;CACd,uBAAuB;CACvB,yBAAyB;AAC1B;;;;;;;AAQA,SAAS,cAAc,IAAqD;CAK3E,MAAM,QAAQ,mDAAmD,KAAK,EAAE;CACxE,IAAI,CAAC,OACJ,OAAO;CAER,OAAO;EACN,OAAO,OAAO,MAAM,EAAE;EACtB,OAAO,MAAM,OAAO,KAAA,IAAY,IAAI,OAAO,MAAM,EAAE;CACpD;AACD;;;;;;;;;;AAWA,SAAgB,yBACf,SACwB;CACxB,MAAM,KAAK,QAAQ,YAAY;CAa/B,IAAI,GAAG,SAAS,QAAQ,GAAG;EAC1B,MAAM,UAAU,cAAc,EAAE;EAChC,MAAM,WACL,YAAY,SACX,QAAQ,QAAQ,KAAM,QAAQ,UAAU,KAAK,QAAQ,SAAS;EAChE,OAAO;GACN,QAAQ;GACR,qBAAqB;GACrB,cAAc;GACd,kCAAkC;GAClC,mBAAmB;GACnB,cAAc;GACd,uBAAuB;GACvB,yBAAyB;GACzB,GAAI,YAAY,OAAO,EAAE,4BAA4B,KAAK,IAAI,CAAC;EAChE;CACD;CAcA,IAAI,GAAG,SAAS,QAAQ,GAAG;EAC1B,MAAM,YAAY,YAAY,EAAE;EAChC,MAAM,kBAAkB,CAAC,mBAAmB,IAAI,SAAS;EACzD,OAAO;GACN,QAAQ;GACR,GAAG;GACH,qBAAqB;GACrB,cAAc;GACd,mBAAmB,CAAC,oBAAoB,IAAI,SAAS;EACtD;CACD;CAUA,IAAI,QAAQ,KAAK,EAAE,GAAG;EACrB,MAAM,kBACL,cAAc,KAAK,EAAE,KAAK,CAAC,iBAAiB,IAAI,YAAY,EAAE,CAAC;EAChE,OAAO;GACN,QAAQ;GACR,qBAAqB;GACrB,cAAc;GACd,kCAAkC;GAClC,mBAAmB;GACnB,cAAc;GACd,uBAAuB;GACvB,yBAAyB;EAC1B;CACD;CAIA,IAAI,YAAY,EAAE,MAAM,eACvB,OAAO;EACN,QAAQ;EACR,qBAAqB;EACrB,cAAc;EACd,kCAAkC;EAClC,mBAAmB;EACnB,cAAc;EACd,uBAAuB;EACvB,yBAAyB;CAC1B;CAID,IAAI,GAAG,SAAS,OAAO,GACtB,OAAO;EACN,QAAQ;EACR,qBAAqB;EACrB,cAAc;EACd,kCAAkC;EAClC,mBAAmB;EACnB,cAAc;EACd,uBAAuB;EACvB,yBAAyB;CAC1B;CAOD,IAAI,GAAG,SAAS,SAAS,GACxB,OAAO;EACN,QAAQ;EACR,GAAG;EACH,mBAAmB;EACnB,cAAc;EACd,uBAAuB;CACxB;CAGD,IAAI,GAAG,SAAS,MAAM,KAAK,GAAG,SAAS,OAAO,GAC7C,OAAO;EACN,QAAQ;EACR,GAAG;EACH,mBAAmB;EACnB,cAAc;CACf;CAGD,IAAI,GAAG,SAAS,KAAK,KAAK,GAAG,SAAS,SAAS,KAAK,GAAG,SAAS,MAAM,GACrE,OAAO;EAAE,QAAQ;EAAS,GAAG;EAAY,mBAAmB;CAAM;CAGnE,OAAO;EAAE,QAAQ;EAAS,GAAG;CAAW;AACzC"}
|
package/dist/package.js
CHANGED
package/package.json
CHANGED
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
{
|
|
2
2
|
"name": "@neon/ai-sdk-provider",
|
|
3
|
-
"version": "0.
|
|
3
|
+
"version": "0.10.0",
|
|
4
4
|
"description": "Community Vercel AI SDK provider for the Neon AI Gateway.",
|
|
5
5
|
"keywords": [
|
|
6
6
|
"neon",
|
|
@@ -45,7 +45,8 @@
|
|
|
45
45
|
"tsdown": "^0.14.1",
|
|
46
46
|
"typescript": "^5.9.0",
|
|
47
47
|
"vitest": "^3.0.9",
|
|
48
|
-
"zod": "^4.4.3"
|
|
48
|
+
"zod": "^4.4.3",
|
|
49
|
+
"@neon/e2e-harness": "0.0.0"
|
|
49
50
|
},
|
|
50
51
|
"engines": {
|
|
51
52
|
"node": ">=20.19.0"
|