ናብ ትሕዝቶ ዝለል
Reasoning effort

Reasoning effort

ሓድሽ

ዝሓስብ ሞዴል ጽሑፍ reasoning ኣብ ጎኒ መልሲ፣ ኣብ ናቱ field ይመልስ። ኣብ ዝተሃንደሱ open-weight ሞዴላት ሓደ field ናይ ጠለብ ሞዴል ቅድሚ ምምላሱ ክሳብ ክንደይ ዕሙቕ ከም ዝሓስብ የዘጋጅእ።

POST https://api.shannon-ai.com/v1/chat/completions 

from openai import OpenAI

client = OpenAI(api_key="YOUR_API_KEY", base_url="https://api.shannon-ai.com/v1")

response = client.chat.completions.create(
    model="Kimi-K3-3BIT-REAP",
    reasoning_effort="low",
    messages=[{"role": "user", "content": "Is 221 a prime number?"}],
)

message = response.choices[0].message
print("Reasoning:", getattr(message, "reasoning_content", None))
print("Answer:", message.content)
print("Reasoning tokens:", response.usage.completion_tokens_details.reasoning_tokens)

እቲ ምላሽ ጽሑፍ reasoning ኣብ reasoning_content፣ እቲ መልሲ ኣብ content፣ ከምኡውን tokens ናይ reasoning ኣብ usage ይሕዝ፦

200 JSON 
{
  "id": "chatcmpl-2a6c8e0b4d1f4a3c9e7b5d3f1a8c6e4b",
  "object": "chat.completion",
  "created": 1791590400,
  "model": "Kimi-K3-3BIT-REAP",
  "choices": [
    {
      "index": 0,
      "message": {
        "role": "assistant",
        "content": "No. 221 = 13 x 17, so it is not prime.",
        "reasoning_content": "Try small primes. 221 / 13 = 17, so 221 = 13 x 17."
      },
      "finish_reason": "stop"
    }
  ],
  "usage": {
    "prompt_tokens": 14,
    "completion_tokens": 43,
    "total_tokens": 57,
    "prompt_tokens_details": {"cached_tokens": 0},
    "completion_tokens_details": {"reasoning_tokens": 27}
  }
}

ኣየኖት ሞዴላት ይሓስቡ

ሞዴል ጽሑፍ reasoning ይመልስ ነቲ effort field ይትግብር
shannon-1.6-lite — —
shannon-1.6-pro, shannon-2-lite, shannon-2-pro, shannon-3, shannon-3-pro, shannon-3.1, shannon-3.1-pro, shannon-coder-1 እወ —
ነፍሲ ወከፍ ዝተሃንደሰ open-weight ሞዴል እወ እወ

እዚኦም flags capabilities.reasoning ከምኡውን capabilities.reasoning_effort እዮም፣ GET /v1/models ንነፍሲ ወከፍ id ዝሕብሮም። ምላሽ ጽሑፍ reasoning ዝሕዝ እቲ ሞዴል ዝፈጠረ እንተ ኾይኑ ጥራይ እዩ። ሞዴላትን ዋጋን

effort ምዕዳል

ነፍሲ ወከፍ endpoint ነቲ effort ናቱ ኣጸሓሕፋ ኣለዎ። ኩሎም ሓደ ካብ ተመሳሳሊ ኣርባዕተ ደረጃታት ይመርጹ።

Field ዓይነት ነባሪ መግለጺ ዝተጠቐመሉ
reasoning_effort string high /v1/chat/completions። እቲ ሞዴል ቅድሚ ምምላሱ ክሳብ ክንደይ ዕሙቕ ከም ዝሓስብ፦ off፣ low፣ medium ወይ high። ዝተሃንደሱ open-weight ሞዴላት
reasoning object | string /v1/chat/completions። ተመሳሳሊ ዋጋ ከም {"effort": "low"} ወይ ከም ጽሩይ string። ክልቲኦም field ታት ምስ ዝለኣኹ፣ reasoning_effort ይቑጸር። ዝተሃንደሱ open-weight ሞዴላት
reasoning.effort string high /v1/responses። ከም object {"effort": "low"} ይጽሓፍ። ዝተሃንደሱ open-weight ሞዴላት
thinking object /v1/messages። {"type": "enabled", "budget_tokens": n} ወይ {"type": "disabled"}። እቲ budget ሓደ ካብ ደረጃታት ይመርጽ፤ ኣብ ታሕቲ ዘሎ ሰንጠረዥ ርአ። ዝተሃንደሱ open-weight ሞዴላት 
{
  "model": "Kimi-K3-3BIT-REAP",
  "reasoning_effort": "low",
  "messages": [{"role": "user", "content": "Is 221 a prime number?"}]
}
ዋጋ ብተወሳኺ ዝቅበል ውጽኢት
off none, minimal Reasoning የለን። እቲ ምላሽ ጽሑፍ reasoning ኣይሕዝን።
low ሕጽር ዝበለ reasoning።
medium default ማእከላይ ዕምቆት።
high max እቲ ዝዓሞቐ ደረጃ። እዚ ጠለብ ዋጋ ዘይልእኽ ደረጃ እዩ።
  • ዋጋታት ብዘይ ኣድናቖት ናብ ዓበይቲ ወይ ንኣሽቲ ፊደላት ይንበቡ።
  • እቲ ዋጋ default ማለት medium እዩ። እቲ field ምድጓይ ማለት high እዩ።
  • እቶም Shannon ሞዴላት (shannon-*) ኣብ ነፍሲ ወከፍ ጠለብ ዕምቆቶም ባዕሎም የዘጋጁ፤ GET /v1/models ንዕኦም reasoning_effort: false ይሕብር።

እቲ budget ናይ ምሕሳብ (thinking) ኣብ መገዲ /v1/messages

ኣብ /v1/messages እቲ field thinking ደረጃ ይመርጽ። budget_tokens ሓደ ካብ ላዕሊ ዘለዉ ደረጃታት ይመርጽ፤ ናቱ ናይ token ወሰን ኣይኮነን። adaptive ከም enabled ይንበብ።

እትልእኾ ደረጃ
{"type": "disabled"} off
{"type": "enabled", "budget_tokens": n} ምስ n ክሳብ 2,048 low
n ካብ 2,049 ክሳብ 8,192 medium
ዝዓበየ n፣ ወይ enabled ብዘይ budget_tokens high
field thinking የለን high

ጽሑፍ reasoning ኣበይ ይበጽሕ

ጽሑፍ reasoning ኩሉ ግዜ ካብ መልሲ ተፈልዩ ይርከብ። ነፍሲ ወከፍ ቅርጺ ናቱ ቦታ ኣለዎ፦

Endpoint ኣይውሕዝን ይውሕዝ
/v1/chat/completions choices[0].message.reasoning_content፦ string፣ ወይ እቲ ሞዴል ዝኾነ ኣይፈጠረን እንተ ኾይኑ null። Chunks ምስ choices[0].delta.reasoning_content፣ ቅድሚ chunks ምስ delta.content።
/v1/responses ሓደ output ኣተዋብ {"type": "reasoning", "summary": [{"type": "summary_text", "text": "…"}]} ቅድሚ እቲ message ኣተዋብ። reasoning እንተ ዘየለ ኣተዋብ የለን። response.output_item.added ምስ reasoning ኣተዋብ፣ ደሓር response.reasoning_summary_text.delta events ከምኡውን ሓደ response.reasoning_summary_text.done።
/v1/messages ቀዳማይ content block {"type": "thinking", "thinking": "…"}፣ ደሓር እቲ text block። reasoning እንተ ዘየለ thinking block የለን። ሓደ thinking content block፦ content_block_start፣ events content_block_delta ምስ thinking_delta፣ ደሓር content_block_stop። እቲ text block ይስዕብ። 
{
  "choices": [
    {
      "index": 0,
      "message": {
        "role": "assistant",
        "content": "No. 221 = 13 x 17, so it is not prime.",
        "reasoning_content": "Try small primes. 221 / 13 = 17, so 221 = 13 x 17."
      },
      "finish_reason": "stop"
    }
  ],
  "usage": {
    "prompt_tokens": 14,
    "completion_tokens": 43,
    "total_tokens": 57,
    "prompt_tokens_details": {"cached_tokens": 0},
    "completion_tokens_details": {"reasoning_tokens": 27}
  }
}

ኣበይ ክርአ፣ ብሞዴል፦

ሞዴላት እቲ ጽሑፍ ኣበይ ይበጽሕ
ዝተሃንደሱ open-weight ሞዴላት ኣብ ነፍሲ ወከፍ endpoint፣ ዝውሕዝ ይኹን ዘይውሕዝ። ጠለብ ምስ effort off፣ ወይ ምስ response_format / text.format፣ ብዘይ ጽሑፍ reasoning ይምለስ።
shannon-3, shannon-3-pro, shannon-3.1, shannon-3.1-pro ኣብ ነፍሲ ወከፍ endpoint፣ ዝውሕዝ ይኹን ዘይውሕዝ፣ እቲ ሞዴል ጽሑፍ reasoning ዝፈጠረ ኩሉ ግዜ።
shannon-2-lite, shannon-2-pro /v1/chat/completions፣ ዘይውሕዝ፦ ኣብ message.reasoning_content። ዝውሕዝ፣ ኣብ ጠለብ ብዘይ tools ወይ response_format፦ እቲ stream እቲ ሞዴል ክሓስብ ከሎ መስመራት ርእይቶ : thinking ይሕዝ፣ ደሓር chunks ናይ መልሲ። /v1/messages፣ ዝውሕዝ፦ ቅድሚ text block ሓደ thinking block።
shannon-1.6-pro, shannon-coder-1 ኣብ ዝውሕዙ ምላሻትን ኣብ ምላሻት ጠለባት tools ዝሕዙን፣ እቲ ሞዴል ጽሑፍ reasoning ዝፈጠረ ኩሉ ግዜ።

ኣብ /v1/responses፣ ዘይውሕዝ ምላሽ ን ስድራ Shannon 3 ከምኡውን ን ዝተሃንደሱ open-weight ሞዴላት ኣተዋብ reasoning ይሕዝ።

stream ምንባብ

ኣብ stream ጽሑፍ reasoning ቀዳማይ ይመጽእ፣ ኣብ ናቱ deltas፣ መልሲ ድማ ይስዕብ። ክልቲኦም ብዝተፈልየ ሕተም ወይ ዕቀብ፦

from openai import OpenAI

client = OpenAI(api_key="YOUR_API_KEY", base_url="https://api.shannon-ai.com/v1")

stream = client.chat.completions.create(
    model="Kimi-K3-3BIT-REAP",
    reasoning_effort="medium",
    stream=True,
    messages=[{"role": "user", "content": "Is 221 a prime number?"}],
)

for chunk in stream:
    if not chunk.choices:
        continue
    delta = chunk.choices[0].delta
    reasoning = getattr(delta, "reasoning_content", None)
    if reasoning:
        print(reasoning, end="", flush=True)      # the reasoning text comes first
    if delta.content:
        print(delta.content, end="", flush=True)  # then the answer

እቲ ቀዳማይ chunk ናይ ዝተሃንደሰ open-weight ሞዴል፣ reasoning ወይ መልሲ፣ እውን delta.role: "assistant" ይሕዝ። ስትሪም

Reasoning tokens ከምኡውን ክፍሊት

ኣብ ዝተሃንደሱ open-weight ሞዴላት፣ reasoning tokens output tokens እዮም፦ ብዋጋ output ናይ ሞዴል ይኽፈሉ፣ ቍጽሪ output ድማ ድሮ የጠቓልሎም። usage ኣብዚ ናይ reasoning ክፍሊ የርኢ፦

Endpoint Field ሓሳብ
/v1/chat/completions usage.completion_tokens_details.reasoning_tokens ድሮ ኣብ completion_tokens ተጠቃሊሉ ኣሎ።
/v1/responses usage.output_tokens_details.reasoning_tokens ድሮ ኣብ output_tokens ተጠቃሊሉ ኣሎ።
/v1/messages usage.output_tokens Reasoning ከምኡውን መልሲ ብሓባር። እዚ ቅርጺ ዝተፈልየ ቍጽሪ የብሉን።
  • ኣብ ዝተሃንደሱ open-weight ሞዴላት max_tokens ጽሑፍ መልሲ የገድብ። Reasoning tokens ኣብኡ ኣይቑጸሩን።
  • reasoning ከም output ስለ ዝኽፈል፣ ደረጃ effort ውን ናይ ዋጋ ምዕዳል እዩ፦ off ዝኾነ reasoning tokens ኣይፈጥርን።
  • እቶም Shannon ሞዴላት prompt_tokens፣ completion_tokens ከምኡውን total_tokens (ኣብ /v1/messages input_tokens ከምኡውን output_tokens) ብዘይ ናይ reasoning ክፍሊ የሕብሩ፣ ን ነፍሲ ወከፍ token ድማ ሓደ ቀዋሚ ዋጋ የኽፍሉ። ሞዴላትን ዋጋን

Reasoning ኣብ ዝቕጽል ጠለብ

reasoning ናብ ሞዴል ክትመልሶ ኣየድልየካን። ኣብ ዝለኣኽካዮ ዕላል ዘሎ thinking block ወይ reasoning ኣተዋብ ናብ ሞዴል ኣይቕጽልን፣ ስለዚ ተራ assistant ብእኡ ወይ ብዘይ እኡ ክምለስ ይከኣል። ሓደ thinking block ክልተ field ታት type ከምኡውን thinking ኣለዎ፤ signature የብሉን።

ጌጋታት

ኩነታት ዓይነት መልእኽቲ መዓስ
400 invalid_request_error unknown reasoning effort '<value>'; expected off, low, medium or high ሓደ ዝተሃንደሰ open-weight ሞዴል ካብ ዝርዝር ወጻኢ ዋጋ ተቐቢሉ፣ ንኣብነት xhigh። ካብ ሚዛንካ ዝኾነ ኣይወሃብን።
422 invalid_request_error Failed to deserialize the JSON body into the target type: … reasoning_effort string ኣይኮነን።

ምርግጋጽ፣ ሚዛንን ጌጋታት server ን ከም ንኩሉ ጠለብ ተመሳሳሊ እዮም። ሓጋይ ኣስተዳደር