update the logic so that all context can fit into smaller context length

[benhaotang]

graph TB
    subgraph Input
    A[Large Context]
    end

    subgraph "Context Processing"
    B[Split Into Chunks]
    C1[Process Chunk 1]
    C2[Process Chunk 2]
    C3[Process Chunk N]
    D[Combine Notes/Comments]
    end

    subgraph "Iterative Refinement"
    E1[Search Planning]
    E2[Writing Planning]
    F1[Generate Queries/Next Steps]
    F2[Refine Writing Plan]
    G1[Final Search Decision]
    G2[Final Report Writing]
    end

    A --> B
    B --> C1
    B --> C2
    B --> C3
    C1 --> D
    C2 --> D
    C3 --> D

    D --> E1
    D --> E2
    E1 --> F1
    E2 --> F2
    F1 --> G1
    F2 --> G2

    style A fill:#f9f,stroke:#333,stroke-width:2px
    style G1 fill:#ccf,stroke:#333,stroke-width:2px
    style G2 fill:#ccf,stroke:#333,stroke-width:2px

    classDef chunk fill:#fcf,stroke:#333,stroke-width:1px
    class C1,C2,C3 chunk
    classDef process fill:#cfc,stroke:#333,stroke-width:1px
    class B,D,E1,E2,F1,F2 process

Loading

Now token number will be estimated and split into chunks, either leave comments per chunk or refine based on chunks

[benhaotang]

Seems only works well with claude sonnet 3.5/3.7, so this PR will only be kept as a draft for now. I think the ultimate solution is to have a memory feature for all the citation for the less capable models.