BofAI · Will-Guan · Apr 8, 2026 · Mar 31, 2026 · Apr 3, 2026 · Apr 3, 2026
diff --git a/docs/llmservice/api/API.md b/docs/llmservice/api/API.md
diff --git a/docs/llmservice/memory.md b/docs/llmservice/memory.md
@@ -0,0 +1,39 @@
+# Memory Service
+
+## Introduction
+
+By default, Large Language Models (LLMs) are stateless—treating every new interaction as a blank slate. The **Memory** service bridges this gap by providing AI agents with long-term continuity. It enables models to retain context, learn user preferences, and deliver a highly personalized experience across multiple sessions, eliminating the need for users to repeat information.
+
+---
+
+## Core Capabilities
+
+### 1. Smart Auto-Learning
+When Memory is active, the system seamlessly evolves through daily interactions. It intelligently identifies and records key user preferences, stylistic choices, and critical facts directly from the conversation stream, building a persistent personalized profile over time.
+
+### 2. External History Import
+To accelerate the personalization process, users can import existing chat histories. By providing exported logs from other applications, the AI can automatically parse, extract, and populate the **Memory Vault** with relevant historical context.
+
+### 3. Absolute User Control
+Transparency and agency are central to our Memory architecture. Through a dedicated management interface, users have full authority to:
+* **Review:** View exactly what information the AI has retained.
+* **Curate:** Manually add custom instructions or specific context.
+* **Manage:** Edit or delete existing memory nodes at any time.
+
+### 4. Incognito Mode
+Privacy is built-in. Users can toggle Memory **OFF** to start an incognito session. In this state, the AI will not retrieve past memories nor record any new data from the ongoing conversation.
+
+---
+
+## Key Advantages
+
+### ⚡ Lightning-Fast & Token-Efficient
+Unlike traditional systems that inject massive history into every prompt, our **Progressive Loading** architecture utilizes a lightweight indexing system. It fetches specific memory nodes **on-demand** only when relevant, ensuring minimal latency and optimal token consumption.
+
+### 🎯 Tailored Companion Experience
+The AI transcends being a generic assistant to become a specialized companion. By remembering project-specific details, formatting rules, and personal workflows, it significantly boosts productivity and interaction quality.
+
+### 🛡️ Zero-Hallucination Recall
+By leveraging exact, user-curated memory nodes rather than broad semantic approximations, the system ensures high-fidelity recall. This prevents the AI from mixing up details or hallucinating past events, providing a "Source of Truth" for user context.
+
+---
diff --git a/docs/llmservice/models/chatgpt-5-2.md → docs/llmservice/models/gpt-5-2.md b/docs/llmservice/models/chatgpt-5-2.md → docs/llmservice/models/gpt-5-2.md
@@ -1,7 +1,7 @@
-# ChatGPT-5.2
+
 
 ## Overview
-ChatGPT-5.2 is the latest generation of the flagship large language model developed by OpenAI. Building upon the powerful capabilities of the 5.1 version, it further optimizes the speed of multimodal processing and the execution efficiency of complex tasks, making it the ideal choice for professional users seeking ultimate performance and efficiency.
+GPT-5.2 is the latest generation of the flagship large language model developed by OpenAI. Building upon the powerful capabilities of the 5.1 version, it further optimizes the speed of multimodal processing and the execution efficiency of complex tasks, making it the ideal choice for professional users seeking ultimate performance and efficiency.
 
 ## Key Features
 * **Efficient Multimodal Processing:** Significantly improves the parsing and generation speed of image and video content compared to 5.1, achieving a smoother multimodal interaction experience.
@@ -27,4 +27,4 @@ ChatGPT-5.2 is the latest generation of the flagship large language model develo
 
 | Model | Input (Credits/Token) | Output (Credits/Token) |
 | :--- | :--- | :--- |
-| **ChatGPT-5.2** | 1.75 | 14.00 |
+| **GPT-5.2** | 1.75 | 14.00 |
diff --git a/docs/llmservice/models/gpt-5-4-mini.md b/docs/llmservice/models/gpt-5-4-mini.md
@@ -0,0 +1,36 @@
+## Overview
+
+GPT-5.4 Mini is OpenAI's high-performance compact model released on March 17, 2026, an efficient distillation of GPT-5.4. It significantly improves over GPT-5 Mini across coding, reasoning, multimodal understanding, and tool use while running 2x+ faster, at roughly 1/6 the cost of the standard model — ideal for high-volume workloads.
+
+## Key Features
+
+- **Near-Flagship Performance**: Scores 54.38% on SWE-Bench Pro, remarkably close to the standard model's 57.7%, at roughly 1/6 the cost.
+- **Strong Scientific Reasoning**: Achieves 87.5% on GPQA Diamond, excelling at graduate-level scientific reasoning tasks.
+- **Full Tool Support**: Supports tool use, web search, image analysis, and Native Computer Use — full capability retention.
+- **2x+ Speed Improvement**: Runs 2x+ faster than GPT-5 Mini, suitable for latency-sensitive workloads.
+- **400K Context Window**: Supports a 400,000 token context window with vision input, suitable for medium-scale long document processing.
+
+## Best Use Cases
+
+- **Coding Assistants & Sub-Agents**: Approaches flagship-level performance on coding benchmarks, delivering reliable code generation and repair at significantly lower cost.
+- **Real-Time AI Applications**: 2x speed improvement makes it ideal for chatbots, real-time translation, and interactive coding assistance.
+- **High-Throughput Data Processing**: Low-cost, high-performance combination suits large-scale document classification, content moderation, and data extraction pipelines.
+- **Desktop Automation Agents**: Full Native Computer Use support for building moderately complex desktop automation workflows.
+
+## Capabilities and Limitations
+
+| Capability             | Detailed Description                                                                                                                   |
+| :--------------------- | :------------------------------------------------------------------------------------------------------------------------------------- |
+| **Reasoning Ability**  | SWE-Bench Pro 54.38%, GPQA Diamond 87.5%; strong reasoning but slightly behind standard on the most complex multi-step problems.       |
+| **Creative Ability**   | Good text and code generation for most everyday creative tasks; less capable than standard for creation requiring very deep reasoning. |
+| **Multimodal Ability** | Supports text and image input with text output; significantly improved multimodal understanding and image analysis over GPT-5 Mini.    |
+| **Response Speed**     | Fast — 2x+ faster than GPT-5 Mini, suitable for latency-sensitive scenarios.                                                           |
+| **Context Window**     | 400,000 tokens                                                                                                                         |
+| **Max Output**         | Not officially specified, estimated 16,000–32,000 tokens                                                                               |
+| **Knowledge Cutoff**   | August 31, 2025                                                                                                                        |
+
+## Credits and Pricing
+
+| Model        | Input (per 1M tokens) | Output (per 1M tokens) |
+| :----------- | --------------------: | ---------------------: |
+| GPT-5.4 Mini |                 $0.75 |                  $4.50 |
diff --git a/docs/llmservice/models/gpt-5-4-nano.md b/docs/llmservice/models/gpt-5-4-nano.md
@@ -0,0 +1,36 @@
+## Overview
+
+GPT-5.4 Nano is the smallest and most cost-effective variant in the GPT-5.4 family, released on March 17, 2026, designed for speed-critical and cost-sensitive scenarios. At just $0.20/$1.25 per MTok, it delivers reasoning capabilities with an Intelligence Index of 44.4, far above the median of 20 among similarly-priced models — ideal for classification, data extraction, ranking, and sub-agent tasks at scale.
+
+## Key Features
+
+- **Extreme Cost Efficiency**: $0.20/1M input + $1.25/1M output, blended rate (3:1 input-to-output ratio) of just $0.46/1M tokens — the most affordable option in the GPT-5.4 family.
+- **High-Speed Inference**: ~221.8 tokens/second generation speed with 3.72-second time to first token, suitable for real-time systems.
+- **Reasoning Model**: Despite being the smallest variant, GPT-5.4 Nano is still a reasoning model with extended thinking / chain-of-thought reasoning.
+- **Multimodal Input**: Supports text and image input, suitable for lightweight multimodal tasks like visual classification and image analysis.
+- **400K Context Window**: Same 400,000 token context window as Mini, providing ample input capacity.
+
+## Best Use Cases
+
+- **Classification & Data Extraction**: OpenAI's officially recommended core use case, delivering reliable performance for structured data processing, text classification, and information extraction.
+- **Coding Sub-Agents**: Suitable for handling simpler supporting tasks in multi-agent architectures, such as code formatting, linting, and small code generation.
+- **Real-Time Systems & High-Throughput Pipelines**: Ultra-low latency and cost combination makes it perfect for background tasks, real-time ranking, and large-scale automation pipelines.
+- **Distributed Agent Architectures**: Serves as edge execution nodes in distributed agent systems, minimizing per-call cost and latency.
+
+## Capabilities and Limitations
+
+| Capability             | Detailed Description                                                                                                                                                 |
+| :--------------------- | :------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| **Reasoning Ability**  | Intelligence Index 44.4 (median 20 for similar price tier), with chain-of-thought reasoning, but less capable than Mini and Standard on complex multi-step problems. |
+| **Creative Ability**   | Suitable for short text generation and template filling; not suited for long-form writing or creation requiring deep reasoning.                                      |
+| **Multimodal Ability** | Supports text and image input with text output; suitable for basic image classification and recognition, complex image analysis is better handled by larger models.  |
+| **Response Speed**     | Very fast — ~221.8 tokens/second, 3.72s time to first token, the fastest model in the GPT-5.4 family.                                                                |
+| **Context Window**     | 400,000 tokens                                                                                                                                                       |
+| **Max Output**         | Not officially specified                                                                                                                                             |
+| **Knowledge Cutoff**   | August 31, 2025                                                                                                                                                      |
+
+## Credits and Pricing
+
+| Model        | Input (per 1M tokens) | Output (per 1M tokens) |
+| :----------- | --------------------: | ---------------------: |
+| GPT-5.4 Nano |                 $0.20 |                   $1.2 |
diff --git a/docs/llmservice/models/gpt-5-4-pro.md b/docs/llmservice/models/gpt-5-4-pro.md
@@ -0,0 +1,44 @@
+## Overview
+
+GPT-5.4 Pro is the highest-performance variant in the GPT-5.4 family, allocating more compute for deeper reasoning to produce smarter and more precise answers on complex, high-stakes tasks. Available exclusively via the Responses API, some requests may take several minutes to complete, making it ideal for accuracy-critical professional use cases.
+
+## Key Features
+
+- **Enhanced Deep Reasoning**: Allocates significantly more reasoning compute, outperforming the standard model on math, science, and complex coding problems.
+- **1M+ Native Context**: Natively supports 1,050,000 token context (922K input + 128K output) without additional configuration.
+- **Unified Capability Inheritance**: Inherits all GPT-5.4 capabilities including Computer Use, Tool Search, and configurable reasoning effort.
+- **Responses API Exclusive**: Available only via the Responses API, specifically optimized for complex tasks requiring deep thinking.
+
+## Best Use Cases
+
+- **High-Stakes Decision Support**: Ideal for financial analysis, legal reasoning, medical diagnostic assistance, and other professional domains requiring extreme accuracy.
+- **Complex Research & Analysis**: The 1M+ context window and deep reasoning make it perfect for cross-document research and systematic analysis.
+- **Advanced Coding Challenges**: Excels on frontier coding benchmarks, suitable for large-scale code refactoring, architecture design, and complex bug diagnosis.
+- **Enterprise AI Agents**: Powers autonomous agent systems requiring the highest reasoning quality, suitable for critical business process automation.
+
+## Capabilities and Limitations
+
+| Capability             | Detailed Description                                                                                                                                                                                                                     |
+| :--------------------- | :--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| **Reasoning Ability**  | Strongest reasoning in the GPT-5.4 family; ranks #1 out of 104 models (overall score 92/100). SWE-bench Pro 57.7%, OSWorld 75% (surpassing 72.4% human expert baseline), GDPval 83%. Configurable reasoning effort: medium, high, xhigh. |
+| **Creative Ability**   | Inherits GPT-5.4's creative capabilities with 128K max output; deeper reasoning improves structured creative tasks like technical writing and system design documents.                                                                   |
+| **Multimodal Ability** | Supports text and image input with text output; image understanding and analysis on par with GPT-5.4 standard.                                                                                                                           |
+| **Tool Use**           | Agentic tool use score 88.9 (#1 of 104 models). Supports Computer Use, Tool Search, and multi-step workflow orchestration.                                                                                                               |
+| **Response Speed**     | Slower — some complex requests may take several minutes, not suitable for low-latency scenarios.                                                                                                                                         |
+| **Context Window**     | 1,050,000 tokens (922K input + 128K output), natively supported without additional configuration.                                                                                                                                        |
+| **Max Output**         | 128,000 tokens                                                                                                                                                                                                                           |
+| **Knowledge Cutoff**   | August 31, 2025                                                                                                                                                                                                                          |
+
+## Credits and Pricing
+
+| Model       | Input (Credits/Token) | Output (Credits/Token) |
+| :---------- | --------------------: | ---------------------: |
+| GPT-5.4 Pro |                 30.00 |                 180.00 |
+
+
+
+## Limitations
+
+- **High latency**: Complex requests may take several minutes; not suitable for real-time or low-latency applications.
+- **Cost**: It is one of the most expensive API models available.
+- **No audio/video input**: Multimodal support is limited to text and image input.
diff --git a/docs/llmservice/models/gpt-5-4.md b/docs/llmservice/models/gpt-5-4.md
@@ -0,0 +1,36 @@
+## Overview
+
+GPT-5.4 is OpenAI's flagship frontier model released on March 5, 2026, the first mainline model to unify reasoning, coding (GPT-5.3-Codex), and computer use into a single architecture. With up to 1,050,000 tokens of context and 128,000 tokens of max output, it is OpenAI's most capable and versatile model to date.
+
+## Key Features
+
+- **Unified Architecture**: Merges reasoning, coding, and computer use into one model — no need to switch between specialized models.
+- **Configurable Reasoning Effort**: Five discrete reasoning levels (none, low, medium, high, xhigh) let developers control thinking depth and cost per query.
+- **Computer Use API**: A new Computer Use API enables the model to see screens, move cursors, click elements, type text, and interact with desktop applications.
+- **Tool Search**: Deferred tool loading mechanism that fetches tool definitions only when needed, reducing total token usage by 47% while maintaining the same accuracy.
+- **1M+ Context Window**: Supports up to 1,050,000 tokens of context (922K input + 128K output), enabling analysis of entire codebases or document collections in a single request.
+
+## Best Use Cases
+
+- **Agentic Coding Assistants**: Scores 57.7% on SWE-Bench Pro, ideal for complex multi-step coding tasks and autonomous code repair.
+- **Desktop Automation & RPA**: OSWorld score of 75% surpasses the human expert baseline of 72.4%, suitable for browser navigation, form filling, and desktop application control.
+- **Knowledge-Intensive Work**: GDPval score of 83% with 33% fewer factual errors per claim than GPT-5.2, ideal for research analysis, document processing, and professional Q&A.
+- **Long-Context Analysis**: The 1M token context window is perfect for legal document review, large-scale code audits, and cross-document correlation analysis.
+
+## Capabilities and Limitations
+
+| Capability | Detailed Description |
+| :--- | :--- |
+| **Reasoning Ability** | SWE-Bench Pro 57.7%, SWE-Bench Verified ~80%, high GPQA Diamond scores, with five configurable reasoning levels. |
+| **Creative Ability** | Excellent long-form text and code generation with 128K max output supporting whole-project generation. |
+| **Multimodal Ability** | Supports text and image input with text output; MMMU Pro score of 81.2%. |
+| **Response Speed** | As a flagship model, moderate inference speed; higher latency in xhigh reasoning mode, near real-time in none mode. |
+| **Context Window** | Standard 272K tokens; expandable to 1,050,000 tokens (requires explicit configuration). |
+| **Max Output** | 128,000 tokens |
+| **Knowledge Cutoff** | August 31, 2025 |
+
+## Credits and Pricing
+
+| Model | Input (per 1M tokens) | Output (per 1M tokens) |
+| :--- | :--- | :--- |
+| GPT-5.4 | $2.50 | $15.00 |
diff --git a/docs/llmservice/models/chatgpt-5-mini.md → docs/llmservice/models/gpt-5-mini.md b/docs/llmservice/models/chatgpt-5-mini.md → docs/llmservice/models/gpt-5-mini.md
@@ -1,7 +1,7 @@
-# ChatGPT-5-mini
+
 
 ## Overview
-ChatGPT-5-mini is an efficient and economical lightweight language model. It is optimized for fast, smooth daily conversations and general tasks, making it a premier choice for cost-effective AI interaction within the Bank of AI ecosystem.
+GPT-5-mini is an efficient and economical lightweight language model. It is optimized for fast, smooth daily conversations and general tasks, making it a premier choice for cost-effective AI interaction within the Bank of AI ecosystem.
 
 ## Key Features
 * **Extremely Fast Response:** Deeply optimized for low response latency, providing a near real-time conversation experience.
@@ -27,4 +27,4 @@ ChatGPT-5-mini is an efficient and economical lightweight language model. It is
 
 | Model | Input (Credits/Token) | Output (Credits/Token) |
 | :--- | :--- | :--- |
-| **ChatGPT-5-mini** | 0.25 | 2.00 |
+| **GPT-5-mini** | 0.25 | 2.00 |