Feature: Agentic AI Workflows for Document Analysis & Automation

[hoangsonww]

Summary

Introduce agentic AI capabilities into DocuThinker so the system can autonomously plan, decide, and execute multi-step tasks on uploaded documents. Instead of only responding to direct user queries, the AI will act as an agent that can break down objectives into sub-tasks (summarization, classification, recommendations, external API queries), execute them, and present structured outcomes.

Problem / Motivation

Currently, DocuThinker processes documents in a single-pass request-response mode (summarization, insights, chat). However:

• Users often need chained workflows (e.g., “Summarize this document, extract actions, send recommendations via email”).
• Complex document sets require multi-step reasoning (e.g., compare insights across multiple PDFs and produce a prioritized report).
• Without agentic planning, the system is reactive instead of proactive.

Adding agentic AI will allow DocuThinker to:

• Autonomously select the right tools (NER, sentiment, summarizer, recommendations).
• Orchestrate multiple steps into workflows.
• Adapt its behavior to user goals instead of static prompts.

---

Proposed Solution

• Integrate a LangChain-style Agent Executor with access to DocuThinker’s internal tools:

  • Summarizer
  • Sentiment Analyzer
  • Document Classifier
  • Recommendations Generator
  • External APIs (GraphQL, REST endpoints)

• Define an Agent Orchestrator Service in backend/ai_ml/agent_executor.js.

• Extend /chat API with an option:

  {
    "query": "Analyze all uploaded documents and generate a decision brief with action items.",
    "mode": "agentic"
  }

• Add a UI workflow visualization (step-by-step execution trace).

---

Architecture

flowchart TD
    User[User Request] --> Intent[Intent Recognition]
    Intent --> AgentPlanner[Agent Planner]
    AgentPlanner -->|Chooses Tool| ToolSummarizer[Summarizer]
    AgentPlanner -->|Chooses Tool| ToolSentiment[Sentiment Analyzer]
    AgentPlanner -->|Chooses Tool| ToolNER[NER & POS Extractor]
    AgentPlanner -->|Chooses Tool| ToolRecommend[Actionable Recs]
    ToolSummarizer --> Memory[Agent Memory]
    ToolSentiment --> Memory
    ToolNER --> Memory
    ToolRecommend --> Memory
    Memory --> Report[Final Structured Response]
    Report --> User

Loading

• Agent Planner: Decides which tools to call and in what order.
• Agent Memory: Maintains context across steps (persistent history).
• Report Generator: Produces final outputs (summary + recs + actions).

---

Acceptance Criteria

• User can select mode=agentic when uploading documents or chatting.
• AI automatically sequences tasks (summarize → extract actions → recommend).
• Execution trace (tools used, steps taken) is visible in UI for transparency.
• Works with multiple documents (cross-doc comparison).
• Agent memory persists within a session for multi-step reasoning.

---

Benefits

• Enables autonomous workflows beyond simple Q&A.
• Adds multi-document intelligence (cross-analysis, comparisons).
• Makes DocuThinker more proactive and adaptive.
• Differentiates it from standard summarization/chat apps.

---

Next Steps

1. Prototype an Agent Orchestrator using LangChain’s initialize_agent with custom tools.
2. Extend backend /chat to support mode=agentic.
3. Add frontend UI component for execution trace visualization.
4. Benchmark with real-world workflows (e.g., meeting notes → action plan).

---