@ailib-official/ai-protocol 0.8.4

package/v1/providers/qwen.yaml

@@ -0,0 +1,192 @@
+$schema: "https://raw.githubusercontent.com/ailib-official/ai-protocol/main/schemas/v1.json"
+
+id: qwen
+protocol_version: "1.5"
+
+name: Qwen (DashScope)
+version: "v1"
+status: stable
+category: ai_provider
+official_url: "https://help.aliyun.com/zh/model-studio"
+support_contact: "https://help.aliyun.com"
+
+endpoint:
+  base_url: "https://dashscope.aliyuncs.com/compatible-mode/v1"
+  protocol: https
+  timeout_ms: 10000
+auth:
+  type: bearer
+  token_env: "QWEN_API_KEY"
+payload_format: "openai_style"
+
+# OpenAI-compatible API families (DashScope compatible-mode)
+api_families: ["chat_completions", "completions"]
+default_api_family: "chat_completions"
+endpoints:
+  # Runtime uses UnifiedRequest.operation (currently "chat") to resolve endpoints.
+  chat:
+    path: "/chat/completions"
+    method: "POST"
+    adapter: "openai"
+  completions:
+    path: "/completions"
+    method: "POST"
+    adapter: "openai"
+
+# Service endpoints (used by online fact-checking / registry verification)
+services:
+  list_models:
+    path: "/models"
+    method: "GET"
+    response_binding: "data"
+
+# Termination reason normalization (see v1/spec.yaml: standard_schema.streaming_events.termination_reasons)
+termination:
+  source_field: "finish_reason"
+  mapping:
+    stop: "end_turn"
+    length: "max_tokens"
+    tool_calls: "tool_use"
+    content_filter: "refusal"
+  notes:
+    - "OpenAI-compatible: finish_reason is reported per candidate."
+
+# Tool invocation normalization (see v1/spec.yaml: standard_schema.content_blocks)
+tooling:
+  source_model: "openai_tool_calls"
+  tool_use:
+    id_path: "id"
+    name_path: "function.name"
+    input_path: "function.arguments"
+    input_format: "json_string"
+  notes:
+    - "OpenAI-compatible tool_calls.function.arguments is typically a JSON string; runtimes SHOULD parse it into an object when normalizing."
+
+# Rate limits (OpenAI-compatible; see OpenAI rate limits guide)
+rate_limit_headers:
+  requests_limit: "x-ratelimit-limit-requests"
+  tokens_limit: "x-ratelimit-limit-tokens"
+  requests_remaining: "x-ratelimit-remaining-requests"
+  tokens_remaining: "x-ratelimit-remaining-tokens"
+  requests_reset: "x-ratelimit-reset-requests"
+  tokens_reset: "x-ratelimit-reset-tokens"
+  retry_after: "retry-after"
+
+# Retry policy (VERIFIED/secondary: short exponential backoff with small retry cap is common in client libs;
+# official docs confirm 429/500/503 meanings but do not prescribe an exact algorithm)
+retry_policy:
+  strategy: "exponential_backoff"
+  max_retries: 2
+  min_delay_ms: 1000
+  max_delay_ms: 8000
+  jitter: "full"
+  retry_on_http_status: [408, 409, 429, 500, 503]
+  notes:
+    - "429 can mean rate limit exceeded (QPS/QPM/other limits) OR quota/payment issue; runtimes SHOULD inspect error messages and avoid infinite retries on quota/payment problems."
+    - "On 429, runtimes SHOULD respect Retry-After when present."
+    - "500/503 are server-side/transient; retry after a brief wait."
+    - "408/409 are commonly retried by client libraries with short exponential backoff."
+
+# Error classification hints (see v1/spec.yaml: standard_schema.error_handling.error_classes)
+error_classification:
+  by_http_status:
+    "400": "invalid_request"
+    "401": "authentication"
+    "429": "rate_limited"
+    "500": "server_error"
+    "503": "overloaded"
+  notes:
+    - "Qwen/DashScope 429 may indicate quota/payment issue; runtimes SHOULD treat as quota_exhausted when the message indicates quota/payment."
+parameter_mappings:
+  temperature: "temperature"
+  max_tokens: "max_tokens"
+  stream: "stream"
+  top_p: "top_p"
+  stop_sequences: "stop"
+  tools: "tools"
+  tool_choice: "tool_choice"
+response_format: "openai_style"
+response_paths:
+  content: "choices[0].message.content"
+  tool_calls: "choices[0].message.tool_calls"
+  usage: "usage"
+  finish_reason: "choices[0].finish_reason"
+streaming:
+  event_format: "data_lines"
+  decoder:
+    format: "sse"
+    delimiter: "\n\n"
+    prefix: "data: "
+    done_signal: "[DONE]"
+  content_path: "choices[0].delta.content"
+  tool_call_path: "choices[0].delta.tool_calls"
+  usage_path: "usage"
+  frame_selector: "exists($.choices) || exists($.error)"
+  event_map:
+    # Text content delta
+    - match: "exists($.choices[*].delta.content)"
+      emit: "PartialContentDelta"
+      fields:
+        content: "$.choices[*].delta.content"
+    # Tool call start
+    - match: "exists($.choices[*].delta.tool_calls[*].function.name)"
+      emit: "ToolCallStarted"
+      fields:
+        tool_call_id: "$.choices[*].delta.tool_calls[*].id"
+        tool_name: "$.choices[*].delta.tool_calls[*].function.name"
+        index: "$.choices[*].delta.tool_calls[*].index"
+    # Tool call arguments streaming
+    - match: "exists($.choices[*].delta.tool_calls[*].function.arguments)"
+      emit: "PartialToolCall"
+      fields:
+        arguments: "$.choices[*].delta.tool_calls[*].function.arguments"
+        index: "$.choices[*].delta.tool_calls[*].index"
+    # Usage metadata
+    - match: "exists($.usage)"
+      emit: "Metadata"
+      fields:
+        usage: "$.usage"
+    # Finish with reason
+    - match: "exists($.choices[*].finish_reason)"
+      emit: "FinalCandidate"
+      fields:
+        finish_reason: "$.choices[*].finish_reason"
+        candidate_index: "$.choices[*].index"
+  stop_condition: "$.choices[0].finish_reason != null"
+features:
+  multi_candidate:
+    support_type: "native"
+    param_name: "n"
+  response_mapping:
+    tool_calls:
+      path: "choices[0].message.tool_calls"
+      fields:
+        id: "id"
+        name: "function.name"
+        args: "function.arguments"
+    error:
+      message_path: "error.message"
+      code_path: "error.code"
+      type_path: "error.type"
+      param_path: "error.param"
+capabilities:
+  streaming: true
+  tools: true
+  vision: true
+  agentic: true
+  parallel_tools: false
+  reasoning: false
+
+availability:
+  required: false
+  regions:
+    - cn
+    - global
+  check:
+    method: GET
+    path: "/models"
+    expected_status: [200, 401]
+    timeout_ms: 3000
+experimental_features:
+  - "multilingual_support"
+  - "qwen_agent"

package/v1/providers/replicate.yaml

@@ -0,0 +1,159 @@
+$schema: "https://raw.githubusercontent.com/ailib-official/ai-protocol/main/schemas/v1.json"
+
+id: replicate
+protocol_version: "1.5"
+
+name: Replicate
+version: "v1"
+status: stable
+category: model_provider
+official_url: "https://replicate.com/docs"
+support_contact: "https://replicate.com/contact"
+
+endpoint:
+  base_url: "https://api.replicate.com/v1"
+  protocol: https
+  timeout_ms: 60000
+
+auth:
+  type: bearer
+  token_env: "REPLICATE_API_TOKEN"
+
+payload_format: "replicate_style"
+
+# Replicate has its own prediction API and OpenAI-compatible endpoints
+api_families: ["predictions", "openai_compatible"]
+default_api_family: "predictions"
+
+endpoints:
+  chat:
+    path: "/models/{owner}/{model}/predictions"
+    method: "POST"
+    adapter: "replicate"
+  # OpenAI-compatible endpoint for supported models
+  openai_chat:
+    path: "/openai/v1/chat/completions"
+    method: "POST"
+    adapter: "openai"
+
+services:
+  list_models:
+    path: "/models"
+    method: "GET"
+    response_binding: "results"
+  get_prediction:
+    path: "/predictions/{id}"
+    method: "GET"
+  cancel_prediction:
+    path: "/predictions/{id}/cancel"
+    method: "POST"
+
+termination:
+  source_field: "status"
+  mapping:
+    succeeded: "end_turn"
+    failed: "other"
+    canceled: "cancelled"
+  notes:
+    - "Replicate uses async predictions with status polling."
+    - "For streaming, use the stream parameter."
+
+tooling:
+  source_model: "unknown"
+  notes:
+    - "Tool calling support varies by model."
+    - "Some models support OpenAI-compatible function calling via the openai endpoint."
+
+retry_policy:
+  strategy: "exponential_backoff"
+  max_retries: 3
+  min_delay_ms: 1000
+  max_delay_ms: 30000
+  jitter: "full"
+  retry_on_http_status: [429, 500, 502, 503, 504]
+  notes:
+    - "Replicate may return 503 when models are cold-starting."
+    - "Consider longer timeouts for initial requests."
+
+error_classification:
+  by_http_status:
+    "400": "invalid_request"
+    "401": "authentication"
+    "402": "quota_exhausted"
+    "403": "permission_denied"
+    "404": "not_found"
+    "422": "invalid_request"
+    "429": "rate_limited"
+    "500": "server_error"
+    "503": "overloaded"
+
+parameter_mappings:
+  temperature: "input.temperature"
+  max_tokens: "input.max_tokens"
+  stream: "stream"
+  top_p: "input.top_p"
+  stop_sequences: "input.stop"
+  prompt: "input.prompt"
+  system_message: "input.system_prompt"
+
+response_format: "replicate_style"
+
+response_paths:
+  content: "output"
+  status: "status"
+  error: "error"
+
+streaming:
+  event_format: "sse"
+  decoder:
+    format: "sse"
+    delimiter: "\n\n"
+    prefix: "data: "
+    done_signal: "done"
+  content_path: "output"
+  event_map:
+    - match: "$.event == 'output'"
+      emit: "PartialContentDelta"
+      fields:
+        content: "$.data"
+    - match: "$.event == 'done'"
+      emit: "StreamEnd"
+      fields:
+        finish_reason: "end_turn"
+    - match: "$.event == 'error'"
+      emit: "StreamError"
+      fields:
+        error: "$.data"
+  stop_condition: "$.event == 'done'"
+
+features:
+  multi_candidate:
+    support_type: "simulated"
+    max_concurrent: 4
+  response_mapping:
+    error:
+      message_path: "detail"
+      type_path: "type"
+
+capabilities:
+  streaming: true
+  tools: false
+  vision: true
+  agentic: false
+  parallel_tools: false
+  reasoning: false
+
+availability:
+  required: false
+  regions:
+    - global
+  check:
+    method: GET
+    path: "/models"
+    expected_status: [200, 401]
+    timeout_ms: 5000
+
+experimental_features:
+  - "async_predictions"
+  - "webhooks"
+  - "cold_boot_optimization"

package/v1/providers/sensenova.yaml

@@ -0,0 +1,153 @@
+$schema: "https://raw.githubusercontent.com/ailib-official/ai-protocol/main/schemas/v1.json"
+
+id: sensenova
+protocol_version: "1.5"
+
+name: 商汤日日新 (SenseNova)
+version: "v1"
+status: stable
+category: ai_provider
+official_url: "https://platform.sensenova.cn/doc"
+support_contact: "https://www.sensetime.com/contact"
+
+endpoint:
+  base_url: "https://api.sensenova.cn/v1"
+  protocol: https
+  timeout_ms: 60000
+
+auth:
+  type: bearer
+  token_env: "SENSENOVA_API_KEY"
+
+payload_format: "openai_style"
+
+# SenseNova uses OpenAI-compatible API
+api_families: ["chat_completions"]
+default_api_family: "chat_completions"
+
+endpoints:
+  chat:
+    path: "/llm/chat-completions"
+    method: "POST"
+    adapter: "sensenova"
+
+services:
+  list_models:
+    path: "/llm/models"
+    method: "GET"
+
+termination:
+  source_field: "finish_reason"
+  mapping:
+    stop: "end_turn"
+    length: "max_tokens"
+    tool_calls: "tool_use"
+  notes:
+    - "日日新 (SenseNova) is SenseTime's flagship AI platform."
+    - "Includes SenseChat and other foundation models."
+
+tooling:
+  source_model: "openai_tool_calls"
+  tool_use:
+    id_path: "id"
+    name_path: "function.name"
+    input_path: "function.arguments"
+    input_format: "json_string"
+  notes:
+    - "Supports function calling similar to OpenAI."
+
+retry_policy:
+  strategy: "exponential_backoff"
+  max_retries: 3
+  min_delay_ms: 1000
+  max_delay_ms: 30000
+  jitter: "full"
+  retry_on_http_status: [429, 500, 502, 503]
+
+error_classification:
+  by_http_status:
+    "400": "invalid_request"
+    "401": "authentication"
+    "403": "permission_denied"
+    "404": "not_found"
+    "429": "rate_limited"
+    "500": "server_error"
+    "503": "overloaded"
+
+parameter_mappings:
+  temperature: "temperature"
+  max_tokens: "max_new_tokens"
+  stream: "stream"
+  top_p: "top_p"
+  frequency_penalty: "repetition_penalty"
+  tools: "tools"
+  tool_choice: "tool_choice"
+
+response_format: "sensenova_style"
+
+response_paths:
+  content: "data.choices[0].message"
+  tool_calls: "data.choices[0].tool_calls"
+  usage: "data.usage"
+  finish_reason: "data.choices[0].finish_reason"
+
+streaming:
+  event_format: "sse"
+  decoder:
+    format: "sse"
+    delimiter: "\n\n"
+    prefix: "data: "
+    done_signal: "[DONE]"
+  content_path: "data.choices[0].delta"
+  usage_path: "data.usage"
+  frame_selector: "exists($.data.choices) || exists($.error)"
+  event_map:
+    - match: "exists($.data.choices[*].delta)"
+      emit: "PartialContentDelta"
+      fields:
+        content: "$.data.choices[*].delta"
+    - match: "exists($.data.choices[*].tool_calls[*].function.name)"
+      emit: "ToolCallStarted"
+      fields:
+        tool_name: "$.data.choices[*].tool_calls[*].function.name"
+    - match: "exists($.data.usage)"
+      emit: "Metadata"
+      fields:
+        usage: "$.data.usage"
+    - match: "exists($.data.choices[*].finish_reason)"
+      emit: "FinalCandidate"
+      fields:
+        finish_reason: "$.data.choices[*].finish_reason"
+  stop_condition: "$.data.choices[0].finish_reason != null"
+
+features:
+  multi_candidate:
+    support_type: "native"
+    param_name: "n"
+  response_mapping:
+    error:
+      message_path: "error.message"
+      code_path: "error.code"
+
+capabilities:
+  streaming: true
+  tools: true
+  vision: true
+  agentic: true
+  parallel_tools: false
+  reasoning: false
+
+availability:
+  required: false
+  regions:
+    - cn
+  check:
+    method: GET
+    path: "/llm/models"
+    expected_status: [200, 401]
+    timeout_ms: 5000
+
+experimental_features:
+  - "multimodal"
+  - "code_generation"
+  - "knowledge_base"

package/v1/providers/siliconflow.yaml

@@ -0,0 +1,156 @@
+$schema: ../../schemas/v1.json
+protocol_version: "1.5"
+id: "siliconflow"
+name: "SiliconFlow"
+status: "stable"
+category: "model_provider"
+official_url: "https://siliconflow.cn/"
+support_contact: "https://siliconflow.cn/support"
+
+endpoint:
+  base_url: "https://api.siliconflow.cn/v1"
+  protocol: "https"
+  timeout_ms: 60000
+
+# Auth + payload format are required for online model fact-checking.
+auth:
+  type: bearer
+  token_env: "SILICONFLOW_API_KEY"
+payload_format: "openai_style"
+
+services:
+  list_models:
+    path: "/models"
+    method: "GET"
+    response_binding: "data"
+
+availability:
+  required: false
+  regions: ["cn", "global"]
+  check:
+    method: "GET"
+    path: "/models"
+    expected_status: [200]
+
+capabilities:
+  streaming: true
+  tools: true
+  vision: true
+  agentic: false
+  reasoning: true
+  parallel_tools: true
+
+# SiliconCloud is OpenAI-compatible
+streaming:
+  event_format: "data_lines"
+  decoder:
+    format: "sse"
+    delimiter: "\n\n"
+    prefix: "data: "
+    done_signal: "[DONE]"
+  frame_selector: "exists($.choices) || exists($.error)"
+  candidate:
+    candidate_id_path: "$.choices[*].index"
+    fan_out: true
+  event_map:
+    - match: "exists($.error)"
+      emit: "StreamError"
+      fields:
+        error: "$.error"
+    - match: "exists($.choices[*].delta.content)"
+      emit: "PartialContentDelta"
+      fields:
+        content: "$.choices[*].delta.content"
+    - match: "exists($.choices[*].delta.tool_calls[*].function.name)"
+      emit: "ToolCallStarted"
+      fields:
+        tool_call_id: "$.choices[*].delta.tool_calls[*].id"
+        tool_name: "$.choices[*].delta.tool_calls[*].function.name"
+        index: "$.choices[*].delta.tool_calls[*].index"
+    - match: "exists($.choices[*].delta.tool_calls[*].function.arguments)"
+      emit: "PartialToolCall"
+      fields:
+        arguments: "$.choices[*].delta.tool_calls[*].function.arguments"
+        index: "$.choices[*].delta.tool_calls[*].index"
+    - match: "exists($.usage)"
+      emit: "Metadata"
+      fields:
+        usage: "$.usage"
+    - match: "exists($.choices[*].finish_reason)"
+      emit: "FinalCandidate"
+      fields:
+        finish_reason: "$.choices[*].finish_reason"
+        candidate_index: "$.choices[*].index"
+  stop_condition: "$.choices[0].finish_reason != null"
+  extra_metadata_path: "$.usage"
+  content_path: "$.choices[0].delta.content"
+  tool_call_path: "$.choices[0].delta.tool_calls"
+
+# Termination reason normalization (SiliconFlow uses OpenAI-compatible finish_reason)
+termination:
+  source_field: "finish_reason"
+  mapping:
+    stop: "end_turn"
+    length: "max_tokens"
+    tool_calls: "tool_use"
+    content_filter: "refusal"
+  notes:
+    - "finish_reason is reported per candidate."
+
+# Tool invocation normalization (SiliconFlow is OpenAI-compatible)
+tooling:
+  source_model: "openai_tool_calls"
+  tool_use:
+    id_path: "id"
+    name_path: "function.name"
+    input_path: "function.arguments"
+    input_format: "json_string"
+  notes:
+    - "SiliconFlow tool_calls.function.arguments is typically a JSON string."
+
+features:
+  multi_candidate:
+    support_type: "native"
+    param_name: "n"
+  response_mapping:
+    tool_calls:
+      path: "choices[0].message.tool_calls"
+      fields:
+        id: "id"
+        name: "function.name"
+        args: "function.arguments"
+    error:
+      message_path: "error.message"
+      code_path: "error.code"
+
+response_paths:
+  content: "choices[0].message.content"
+  tool_calls: "choices[0].message.tool_calls"
+  usage: "usage"
+  finish_reason: "choices[0].finish_reason"
+
+parameter_mappings:
+  temperature: "temperature"
+  max_tokens: "max_tokens"
+  stream: "stream"
+  top_p: "top_p"
+  stop_sequences: "stop"
+  tools: "tools"
+  tool_choice: "tool_choice"
+
+retry_policy:
+  max_retries: 3
+  max_delay_ms: 30000
+  strategy: "exponential_backoff"
+  min_delay_ms: 1000
+  jitter: "full"
+  retry_on_http_status: [429, 500, 502, 503]
+
+error_classification:
+  by_http_status:
+    "400": "invalid_request"
+    "401": "authentication"
+    "403": "permission_denied"
+    "429": "rate_limited"
+    "500": "server_error"
+    "503": "overloaded"