Documentation
¶
Overview ¶
Package chatcompletions is a transport-agnostic codec for the OpenAI Chat Completions wire format (`/chat/completions`-style request/response and SSE streaming). It maps between contenox's neutral modelrepo types and the OpenAI-compatible JSON shape, performing tool-name sanitization and round-tripping.
It does NO I/O: callers build a Request, marshal and POST it through their own transport (API-key header for direct OpenAI/Mistral), then hand the raw response bytes back here to decode. This is what lets the direct Mistral / OpenAI provider stay a thin transport wrapper around the shared codec.
Index ¶
Constants ¶
This section is empty.
Variables ¶
This section is empty.
Functions ¶
func DecodeResponse ¶
DecodeResponse parses a non-streaming response into a neutral ChatResult, translating sanitized tool-call names back via nameMap.
Types ¶
type Request ¶
type Request struct {
Model string `json:"model"`
Messages []wireMessage `json:"messages"`
Temperature *float64 `json:"temperature,omitempty"`
MaxTokens *int `json:"max_tokens,omitempty"`
TopP *float64 `json:"top_p,omitempty"`
Seed *int `json:"seed,omitempty"`
Tools []wireTool `json:"tools,omitempty"`
ToolChoice string `json:"tool_choice,omitempty"`
Stream bool `json:"stream,omitempty"`
}
Request is the OpenAI-compatible chat/completions request body.
Note: this codec emits `max_tokens` (the field Mistral's OpenAI-compatible endpoint accepts), not the newer `max_completion_tokens`.
func Build ¶
func Build(model string, messages []modelrepo.Message, cfg *modelrepo.ChatConfig) (Request, map[string]string)
Build converts neutral messages + config into a chat/completions Request. It returns a nameMap (sanitized tool name -> original) so DecodeResponse / the StreamDecoder can translate tool-call names back to what the caller used.
model is placed verbatim in the body; the transport decides the exact string (e.g. "mistral-large-latest", or whatever id the provider expects).
type Response ¶
type Response struct {
Choices []struct {
Index int `json:"index"`
Message responseMsg `json:"message"`
FinishReason string `json:"finish_reason"`
} `json:"choices"`
}
Response is the non-streaming chat/completions response body.
type StreamDecoder ¶
type StreamDecoder struct {
// contains filtered or unexported fields
}
StreamDecoder assembles a streamed chat/completions response. Text and reasoning deltas are returned as parcels; tool-call fragments are accumulated per index and exposed via ToolCalls() once the stream ends.
func NewStreamDecoder ¶
func NewStreamDecoder(nameMap map[string]string) *StreamDecoder
NewStreamDecoder returns a decoder. nameMap is the sanitized->original map from Build (may be nil if no tools).
func (*StreamDecoder) DecodeLine ¶
func (d *StreamDecoder) DecodeLine(payload []byte) (*modelrepo.StreamParcel, error)
DecodeLine parses one SSE data payload (the bytes AFTER the "data: " prefix, excluding the "[DONE]" sentinel which the caller should skip). It returns a parcel to emit if the chunk carried visible text/reasoning, or nil.
func (*StreamDecoder) ToolCalls ¶
func (d *StreamDecoder) ToolCalls() []modelrepo.ToolCall
ToolCalls returns the tool calls assembled across the stream, in index order, with names translated back via nameMap.
Source Files