mirror of
https://github.com/hanzoai/ai.git
synced 2026-11-20 14:13:23 +00:00
DeepSeek upstreams (zen5-pro→deepseek-v4-pro, zen5-flash→deepseek-4-flash)
inline chain-of-thought into `content` as `…reasoning…</think><answer>` — the
opening <think> is consumed by the chat template, so only the closing </think>
survives. GLM upstreams (zen5, zen5-coder) instead use the separate
reasoning_content field. The relay forwarded DeepSeek content verbatim, so its
</think> template token leaked into the visible answer and broke the agent's
character in agent and the console.
One normalizer, one place: model.{InlinesReasoning,StripLeadingReasoning,
ReasoningStripper} is the single source of truth (package model, so model/openai.go
can use it without a cycle). Every relay surface strips only when the upstream
inlines reasoning; every other stream is byte-for-byte unchanged.
- model/reasoning.go — core: gate + whole-string strip + streaming stripper
(64 KiB cap so a missing </think> can never buffer unboundedly).
- controllers/reasoning.go — OpenAI-SSE wrappers (applyReasoningStrip /
stripReasoningBody), fail-safe passthrough on any parse error.
- controllers/anthropic_translate.go — the agent path (Anthropic in →
OpenAI upstream → Anthropic out): strip in both the streaming translator and
openAIResponseToAnthropic. Renamed the shadowing `model` param → `modelID` to
remove the package-shadow footgun (values, not places).
- controllers/openai_api.go, billing_reserve.go, model/openai.go (QueryText) —
wire the stripper on every remaining forward path. Billing always tokenizes
the ORIGINAL (reasoning still counted); only the forwarded text is cleaned.
Tests: streaming strip across a chunk-split </think>, non-stream strip, GLM
reasoning_content→thinking preserved, non-DeepSeek content untouched (gate),
billing unchanged. TDD green.
88 lines
3.6 KiB
Go
88 lines
3.6 KiB
Go
// Copyright 2023-2025 Hanzo AI Inc. All Rights Reserved.
|
|
//
|
|
// Licensed under the Apache License, Version 2.0 (the "License");
|
|
// you may not use this file except in compliance with the License.
|
|
// You may obtain a copy of the License at
|
|
//
|
|
// http://www.apache.org/licenses/LICENSE-2.0
|
|
//
|
|
// Unless required by applicable law or agreed to in writing, software
|
|
// distributed under the License is distributed on an "AS IS" BASIS,
|
|
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
// See the License for the specific language governing permissions and
|
|
// limitations under the License.
|
|
|
|
package model
|
|
|
|
import "strings"
|
|
|
|
// Reasoning normalization across heterogeneous upstreams.
|
|
//
|
|
// Zen aliases front models that expose chain-of-thought differently: GLM (zen5,
|
|
// zen5-coder) returns it in the separate reasoning_content field, but DeepSeek
|
|
// (zen5-pro → deepseek-v4-pro, zen5-flash → deepseek-4-flash) inlines it into
|
|
// content as `<reasoning></think><answer>` — the opening <think> consumed by its
|
|
// chat template. Presented through one Anthropic/OpenAI-shaped surface with no
|
|
// normalization, DeepSeek's closing </think> and its reasoning text leak into the
|
|
// visible answer. These helpers strip that leading block so every upstream reads
|
|
// identically: clean content, no template artifacts. Lives in package model so
|
|
// both the relay (controllers) and the QueryText providers share one definition.
|
|
|
|
const thinkClose = "</think>"
|
|
|
|
// reasoningCap bounds how much leading content is held while waiting for the
|
|
// </think> terminator, so a stream gated as reasoning-inlining but somehow
|
|
// lacking the terminator flushes rather than buffering unboundedly.
|
|
const reasoningCap = 1 << 16 // 64 KiB
|
|
|
|
// InlinesReasoning reports whether an upstream model emits chain-of-thought
|
|
// inline in content (terminated by </think>) rather than in reasoning_content.
|
|
// Gating on it keeps every other upstream's bytes untouched — no buffering, no
|
|
// risk of a false strip.
|
|
func InlinesReasoning(upstreamModel string) bool {
|
|
return strings.Contains(strings.ToLower(upstreamModel), "deepseek")
|
|
}
|
|
|
|
// StripLeadingReasoning removes a leading `<think>…</think>` (or bare `…</think>`,
|
|
// opener consumed by the template) block from a complete content string,
|
|
// returning the clean answer. Content with no </think> is returned unchanged —
|
|
// a no-op on already-clean (e.g. GLM) output.
|
|
func StripLeadingReasoning(content string) string {
|
|
if i := strings.Index(content, thinkClose); i >= 0 {
|
|
return strings.TrimLeft(content[i+len(thinkClose):], " \n")
|
|
}
|
|
return content
|
|
}
|
|
|
|
// ReasoningStripper strips the leading inline-reasoning block from a STREAMED
|
|
// content sequence. It holds content deltas until </think> arrives, then emits
|
|
// the answer remainder and passes every later delta straight through. Feed it
|
|
// only content deltas from a reasoning-inlining upstream (see InlinesReasoning);
|
|
// the terminator may split across deltas, which the running buffer handles.
|
|
type ReasoningStripper struct {
|
|
done bool
|
|
buf strings.Builder
|
|
}
|
|
|
|
// Feed returns the content to forward for one upstream content delta: "" while
|
|
// still inside the reasoning block, the answer text once </think> is seen (or the
|
|
// held buffer if the cap is hit without a terminator), then passthrough.
|
|
func (s *ReasoningStripper) Feed(delta string) string {
|
|
if s.done {
|
|
return delta
|
|
}
|
|
s.buf.WriteString(delta)
|
|
b := s.buf.String()
|
|
if i := strings.Index(b, thinkClose); i >= 0 {
|
|
s.done = true
|
|
s.buf.Reset()
|
|
return strings.TrimLeft(b[i+len(thinkClose):], " \n")
|
|
}
|
|
if s.buf.Len() >= reasoningCap {
|
|
s.done = true
|
|
s.buf.Reset()
|
|
return b
|
|
}
|
|
return ""
|
|
}
|