zack3d/big-agi
1
0
Fork 0
mirror of https://github.com/enricoros/big-AGI.git synced 2026-05-10 21:50:14 -07:00
Code Issues Packages Projects Releases Wiki Activity

AIX: DeepSeek V4: fix function calling

Browse Source
This commit is contained in:
Enrico Ros
2026-04-24 05:38:41 -07:00
parent 9f89243d7f
commit cff3d90613
2 changed files with 36 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/modules/aix/client/aix.client.chatGenerateRequest.ts
+7 -3
View File
@@ -409,11 +409,15 @@ export async function aixCGR_ChatSequence_FromDMessagesOrThrow(
break;
case 'ma':
// Preserve reasoning continuity across turns. Two channels, any one is sufficient:
// Preserve reasoning continuity across turns. Three channels, any one is sufficient:
// - Anthropic: part.textSignature / part.redactedData (bespoke fields, see Anthropic extended thinking docs)
// - OpenAI/Gemini: _vnd sidecar (reasoningItem.* / thoughtSignature, generic vendor-state mechanism)
// - OpenAI Responses / Gemini: _vnd sidecar (reasoningItem.* / thoughtSignature, opaque continuity handle)
// - DeepSeek V4 (OpenAI chat-completions): plain reasoning text in aText is the payload itself
const oaiReasoning = _vnd?.openai?.reasoningItem;
const hasReasoningHandle = aPart.textSignature || aPart.redactedData?.length || oaiReasoning?.encryptedContent || oaiReasoning?.id;
const hasReasoningHandle =
(aPart.textSignature || aPart.redactedData?.length)
|| (oaiReasoning?.encryptedContent || oaiReasoning?.id)
|| (aPart.aText && aPart.aType === 'reasoning'); // DeepSeek V4 reasoning in plain text - NOTE: will send LOTS of 'ma' parts (e.g. to Gemini, which doesn't even need them)
if (hasReasoningHandle) {
const aModelAuxPart = aPart as AixParts_ModelAuxPart; // NOTE: this is a forced cast from readonly string[] to string[], but not a big deal here
modelMessage.parts.push(_vnd ? { ...aModelAuxPart, _vnd } : aModelAuxPart);
src/modules/aix/server/dispatch/chatGenerate/adapters/openai.chatCompletions.ts
+29 -3
View File
@@ -60,7 +60,7 @@ export function aixToOpenAIChatCompletions(openAIDialect: OpenAIDialects, model:
throw new Error('This service does not support function calls');
// Convert the chat messages to the OpenAI 4-Messages format
let chatMessages = _toOpenAIMessages(chatGenerate.systemMessage, chatGenerate.chatSequence, hotFixOpenAIOFamily);
let chatMessages = _toOpenAIMessages(openAIDialect, chatGenerate.systemMessage, chatGenerate.chatSequence, hotFixOpenAIOFamily);
// Apply hotfixes
@@ -70,6 +70,13 @@ export function aixToOpenAIChatCompletions(openAIDialect: OpenAIDialects, model:
if (hotFixAlternateUserAssistantRoles)
chatMessages = _fixAlternateUserAssistantRoles(chatMessages);
// [DeepSeek, 2026-04-24] When tools are present and thinking isn't disabled, V4 demands reasoning_content on EVERY assistant message in history
// Inject '' placeholder where missing; real reasoning is attached by _toOpenAIMessages
if (openAIDialect === 'deepseek' && chatGenerate.tools?.length)
for (const m of chatMessages)
if (m.role === 'assistant' && m.reasoning_content === undefined)
m.reasoning_content = '';
// constrained output modes - both JSON and tool invocations
// const strictJsonOutput = !!model.strictJsonOutput;
@@ -452,7 +459,10 @@ function _fixVndOaiRestoreMarkdown_Inline(payload: TRequest) {
}*/
function _toOpenAIMessages(systemMessage: AixMessages_SystemMessage | null, chatSequence: AixMessages_ChatMessage[], hotFixOpenAIo1Family: boolean): TRequestMessages {
function _toOpenAIMessages(openAIDialect: OpenAIDialects, systemMessage: AixMessages_SystemMessage | null, chatSequence: AixMessages_ChatMessage[], hotFixOpenAIo1Family: boolean): TRequestMessages {
// [DeepSeek, 2026-04-24] V4 thinking-by-default - reasoning_content must round-trip on tool-call turns; payload is the 'ma' part's aText (unlike Gemini/OpenAI-Responses which carry opaque handles).
const echoDeepseekReasoning = openAIDialect === 'deepseek';
// Transform the chat messages into OpenAI's format (an array of 'system', 'user', 'assistant', and 'tool' messages)
const chatMessages: TRequestMessages = [];
@@ -565,6 +575,8 @@ function _toOpenAIMessages(systemMessage: AixMessages_SystemMessage | null, chat
break;
case 'model':
// Accumulate 'ma' reasoning text across this turn; echoed below onto the assistant message if it carries tool_calls (DeepSeek only).
let pendingReasoningText = '';
for (const part of parts) {
const currentMessage = chatMessages[chatMessages.length - 1];
switch (part.pt) {
@@ -640,7 +652,9 @@ function _toOpenAIMessages(systemMessage: AixMessages_SystemMessage | null, chat
break;
case 'ma':
// ignore this thinking block - Anthropic only
// [DeepSeek only] accumulate reasoning text for the echo-back below. Other dialects ignore 'ma' (reasoning continuity flows via _vnd opaque handles, not via this adapter).
if (echoDeepseekReasoning && part.aType === 'reasoning' && part.aText)
pendingReasoningText += part.aText;
break;
case 'tool_response':
@@ -661,6 +675,18 @@ function _toOpenAIMessages(systemMessage: AixMessages_SystemMessage | null, chat
}
}
// [DeepSeek] attach accumulated reasoning to this turn's assistant message only if it carries tool_calls; plain-text turns don't need the echo per docs.
if (echoDeepseekReasoning && pendingReasoningText) {
for (let i = chatMessages.length - 1; i >= 0; i--) {
const m = chatMessages[i];
if (m.role !== 'assistant') continue;
if (m.tool_calls?.length)
m.reasoning_content = pendingReasoningText;
break; // stop at the most recent assistant message from this turn
}
}
break;
}
}