Chat with PDF Inputs

post

Header parameters

AuthorizationstringRequiredExample: sk-Yf4DN

Body

modelstringRequired

Responses

200Success

application/json

idstringRequired

modelstringRequired

objectstringRequired

createdintegerRequired

request_idstringRequired

post

/v1/chat/completions

POST /v1/chat/completions HTTP/1.1
Host: llm.onerouter.pro
Authorization: text
Content-Type: application/json
Accept: */*
Content-Length: 2415

{
  "model": "openai/gpt-5.4",
  "messages": [
    {
      "role": "system",
      "content": "You are a highly capable AI assistant with access to tools. Respond concisely and accurately."
    },
    {
      "role": "user",
      "content": [
        {
          "type": "text",
          "text": "What are the main points in this document?"
        },
        {
          "type": "file",
          "file": {
            "filename": "gpt-4-technical-report.pdf",
            "file_data": "data:application/pdf;base64,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"
          }
        }
      ]
    }
  ]
}

200Success

{
  "id": "chatcmpl-DKzDUSxMFeRIgXVE9ZQd3QV5Ao8hd",
  "model": "openai/gpt-5.4",
  "object": "chat.completion",
  "created": 1773894248,
  "choices": [
    {
      "index": 0,
      "message": {
        "role": "assistant",
        "content": "Here are the main points of the document, the **GPT-4 Technical Report**:\n\n### 1. GPT-4 is a multimodal model\n- GPT-4 can take **text and image inputs** and produce **text outputs**.\n- It is based on a **Transformer** architecture and trained primarily by **next-token prediction**, then improved with **RLHF** (reinforcement learning from human feedback).\n\n### 2. Strong performance on exams and benchmarks\n- The report says GPT-4 achieves **human-level performance on several professional and academic exams**.\n- Notable example: on a **simulated bar exam**, GPT-4 scores around the **top 10%**, while GPT-3.5 was around the bottom 10% (pages 1, 5–6).\n- On standard NLP benchmarks, GPT-4 outperforms prior large language models on tasks like:\n  - **MMLU**\n  - **HellaSwag**\n  - **ARC**\n  - **HumanEval**\n  - **GSM-8K**\n- The **table on page 7** shows GPT-4 beating previous few-shot language model results on nearly every listed benchmark.\n\n### 3. Predictable scaling was a major technical achievement\n- A key focus was building infrastructure so performance would scale **predictably**.\n- OpenAI says they could **forecast GPT-4’s final loss and some capabilities** using models trained with **1,000x–10,000x less compute** (pages 2–4).\n- The charts on **pages 3–4** show these predictions matched actual GPT-4 results closely.\n\n### 4. Strong multilingual capability\n- GPT-4 performs well beyond English.\n- On translated **MMLU** tests, it exceeded the English-language state of the art in **24 of 26 languages** tested (pages 1, 8).\n- The **chart on page 8** shows strong results across many languages, including lower-resource ones like **Latvian, Welsh, and Swahili**.\n\n### 5. Visual input capability\n- GPT-4 can interpret images combined with text prompts.\n- Examples on **pages 9 and 34–39** show it:\n  - explaining jokes in images,\n  - reading charts,\n  - solving diagram-based questions,\n  - summarizing paper screenshots.\n\n### 6. Important limitations remain\n- GPT-4 still:\n  - **hallucinates**,\n  - makes reasoning mistakes,\n  - has a **limited context window**,\n  - does **not learn from experience**,\n  - has knowledge that is mostly cut off around **September 2021** (pages 1, 10).\n- The report emphasizes it is **not fully reliable**, especially in high-stakes settings.\n\n### 7. Factuality improved, but not solved\n- GPT-4 performs better than GPT-3.5 on internal factuality evaluations.\n- On **page 10**, OpenAI reports a **19 percentage point improvement** over its latest GPT-3.5 model on internal adversarial factuality tests.\n- It also improves on **TruthfulQA** (pages 10–11), though errors remain.\n\n### 8. Safety and risk mitigation were a major focus\n- The report highlights risks involving:\n  - **bias**\n  - **disinformation**\n  - **privacy**\n  - **cybersecurity**\n  - **weaponization/proliferation**\n  - **overreliance**\n  - **emergent risky behavior**\n- OpenAI used:\n  - **expert red teaming** with over 50 specialists,\n  - additional **safety-focused RLHF** data,\n  - **rule-based reward models**,\n  - monitoring and policy systems (pages 11–14, and the System Card starting page 41).\n\n### 9. Safety improved substantially over earlier models\n- OpenAI says GPT-4 is much better at refusing unsafe requests.\n- On **page 13**, they report:\n  - an **82% reduction** in responding to disallowed requests vs GPT-3.5,\n  - **29% better compliance** with policy on sensitive requests,\n  - much lower toxic output rates on **RealToxicityPrompts**.\n\n### 10. But safety is still brittle\n- The report is clear that **jailbreaks still exist**.\n- Even with safety tuning, users can sometimes elicit disallowed behavior using adversarial prompts or system-message attacks (pages 14, 68–69).\n- OpenAI says deployment-time controls and ongoing monitoring are still necessary.\n\n### 11. Limited transparency on some training details\n- OpenAI intentionally **does not disclose** model size, architecture details, hardware, training compute, or full dataset details (page 2).\n- They say this is due to **competitive and safety considerations**.\n\n### 12. Overall conclusion\n- GPT-4 is presented as a **significant step forward** in capability, multimodality, multilingual performance, and alignment.\n- But the report repeatedly stresses that it also introduces **new societal and safety challenges**, and that more work is needed before such systems are broadly safe and reliable (page 14).\n\nIf you want, I can also give you:\n1. a **5-bullet ultra-short summary**, or  \n2. a **section-by-section summary** of the whole report."
      },
      "finish_reason": "stop",
      "logprobs": null
    }
  ],
  "request_id": "7427fde627954329bf2d2c4a0e17b610",
  "usage": {
    "prompt_tokens": 110924,
    "completion_tokens": 1116,
    "total_tokens": 112040,
    "prompt_tokens_details": {
      "cached_tokens": 0,
      "cache_write_tokens": 0,
      "audio_tokens": 0,
      "video_tokens": 0
    },
    "completion_tokens_details": {
      "reasoning_tokens": 0,
      "image_tokens": 0,
      "audio_tokens": 0
    }
  }
}

PreviousChat with Images Inputs NextChat with Tool Calling

Last updated 1 month ago

POST /v1/chat/completions HTTP/1.1
Host: llm.onerouter.pro
Authorization: text
Content-Type: application/json
Accept: */*
Content-Length: 2415

{
  "model": "openai/gpt-5.4",
  "messages": [
    {
      "role": "system",
      "content": "You are a highly capable AI assistant with access to tools. Respond concisely and accurately."
    },
    {
      "role": "user",
      "content": [
        {
          "type": "text",
          "text": "What are the main points in this document?"
        },
        {
          "type": "file",
          "file": {
            "filename": "gpt-4-technical-report.pdf",
            "file_data": "data:application/pdf;base64,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"
          }
        }
      ]
    }
  ]
}

{
  "id": "chatcmpl-DKzDUSxMFeRIgXVE9ZQd3QV5Ao8hd",
  "model": "openai/gpt-5.4",
  "object": "chat.completion",
  "created": 1773894248,
  "choices": [
    {
      "index": 0,
      "message": {
        "role": "assistant",
        "content": "Here are the main points of the document, the **GPT-4 Technical Report**:\n\n### 1. GPT-4 is a multimodal model\n- GPT-4 can take **text and image inputs** and produce **text outputs**.\n- It is based on a **Transformer** architecture and trained primarily by **next-token prediction**, then improved with **RLHF** (reinforcement learning from human feedback).\n\n### 2. Strong performance on exams and benchmarks\n- The report says GPT-4 achieves **human-level performance on several professional and academic exams**.\n- Notable example: on a **simulated bar exam**, GPT-4 scores around the **top 10%**, while GPT-3.5 was around the bottom 10% (pages 1, 5–6).\n- On standard NLP benchmarks, GPT-4 outperforms prior large language models on tasks like:\n  - **MMLU**\n  - **HellaSwag**\n  - **ARC**\n  - **HumanEval**\n  - **GSM-8K**\n- The **table on page 7** shows GPT-4 beating previous few-shot language model results on nearly every listed benchmark.\n\n### 3. Predictable scaling was a major technical achievement\n- A key focus was building infrastructure so performance would scale **predictably**.\n- OpenAI says they could **forecast GPT-4’s final loss and some capabilities** using models trained with **1,000x–10,000x less compute** (pages 2–4).\n- The charts on **pages 3–4** show these predictions matched actual GPT-4 results closely.\n\n### 4. Strong multilingual capability\n- GPT-4 performs well beyond English.\n- On translated **MMLU** tests, it exceeded the English-language state of the art in **24 of 26 languages** tested (pages 1, 8).\n- The **chart on page 8** shows strong results across many languages, including lower-resource ones like **Latvian, Welsh, and Swahili**.\n\n### 5. Visual input capability\n- GPT-4 can interpret images combined with text prompts.\n- Examples on **pages 9 and 34–39** show it:\n  - explaining jokes in images,\n  - reading charts,\n  - solving diagram-based questions,\n  - summarizing paper screenshots.\n\n### 6. Important limitations remain\n- GPT-4 still:\n  - **hallucinates**,\n  - makes reasoning mistakes,\n  - has a **limited context window**,\n  - does **not learn from experience**,\n  - has knowledge that is mostly cut off around **September 2021** (pages 1, 10).\n- The report emphasizes it is **not fully reliable**, especially in high-stakes settings.\n\n### 7. Factuality improved, but not solved\n- GPT-4 performs better than GPT-3.5 on internal factuality evaluations.\n- On **page 10**, OpenAI reports a **19 percentage point improvement** over its latest GPT-3.5 model on internal adversarial factuality tests.\n- It also improves on **TruthfulQA** (pages 10–11), though errors remain.\n\n### 8. Safety and risk mitigation were a major focus\n- The report highlights risks involving:\n  - **bias**\n  - **disinformation**\n  - **privacy**\n  - **cybersecurity**\n  - **weaponization/proliferation**\n  - **overreliance**\n  - **emergent risky behavior**\n- OpenAI used:\n  - **expert red teaming** with over 50 specialists,\n  - additional **safety-focused RLHF** data,\n  - **rule-based reward models**,\n  - monitoring and policy systems (pages 11–14, and the System Card starting page 41).\n\n### 9. Safety improved substantially over earlier models\n- OpenAI says GPT-4 is much better at refusing unsafe requests.\n- On **page 13**, they report:\n  - an **82% reduction** in responding to disallowed requests vs GPT-3.5,\n  - **29% better compliance** with policy on sensitive requests,\n  - much lower toxic output rates on **RealToxicityPrompts**.\n\n### 10. But safety is still brittle\n- The report is clear that **jailbreaks still exist**.\n- Even with safety tuning, users can sometimes elicit disallowed behavior using adversarial prompts or system-message attacks (pages 14, 68–69).\n- OpenAI says deployment-time controls and ongoing monitoring are still necessary.\n\n### 11. Limited transparency on some training details\n- OpenAI intentionally **does not disclose** model size, architecture details, hardware, training compute, or full dataset details (page 2).\n- They say this is due to **competitive and safety considerations**.\n\n### 12. Overall conclusion\n- GPT-4 is presented as a **significant step forward** in capability, multimodality, multilingual performance, and alignment.\n- But the report repeatedly stresses that it also introduces **new societal and safety challenges**, and that more work is needed before such systems are broadly safe and reliable (page 14).\n\nIf you want, I can also give you:\n1. a **5-bullet ultra-short summary**, or  \n2. a **section-by-section summary** of the whole report."
      },
      "finish_reason": "stop",
      "logprobs": null
    }
  ],
  "request_id": "7427fde627954329bf2d2c4a0e17b610",
  "usage": {
    "prompt_tokens": 110924,
    "completion_tokens": 1116,
    "total_tokens": 112040,
    "prompt_tokens_details": {
      "cached_tokens": 0,
      "cache_write_tokens": 0,
      "audio_tokens": 0,
      "video_tokens": 0
    },
    "completion_tokens_details": {
      "reasoning_tokens": 0,
      "image_tokens": 0,
      "audio_tokens": 0
    }
  }
}

hashtagChat with PDF Inputs

Chat with PDF Inputs