Создаем AI-агентов на LangChain/LangGraph от идеи до MVP за 20 минут!

Summary of Technological Concepts and What the Tutorial Builds

What “agents” are (conceptual overview)

• Agents are systems where multiple models/components (often LLMs) exchange information, make decisions, and produce a result.
• This tutorial focuses on LLM-based agents, but the underlying idea can generalize to other generator systems (e.g., image/video/presentation generation).
• A common framing is that an agent is coordinated as a graph or chain of steps.

Why LangGraph-style orchestration matters

• The agent is implemented as a graph:
  • Nodes represent function/model actions.
  • Edges pass data to subsequent nodes.
• This approach is presented as a response to LLM context limits—for example, handling huge documents via hierarchical summarization:
  1. LLM A summarizes each report.
  2. LLM B summarizes the combined intermediate summaries.
• The same “divide-and-control” idea extends to multi-step tasks that may loop, request clarification, or search additional sources.

---

The Implemented Product/Tutorial Agent (Core Functionality)

Goal of the tutorial: a “research agent”

The agent workflow:

1. Takes a broad research topic.
2. Asks an LLM to propose three narrower subtopics.
3. Requests user approval (yes/no) for the proposed subtopics.
4. If not approved, the agent regenerates subtopics (loop).
5. After approval, it:
   • searches an archive,
   • collects relevant documents,
   • analyzes them,
   • identifies research gaps, and
   • generates a final report.

Automated workflow logic (decision loop)

• The LangGraph keeps a shared state object containing:
  • the topic
  • the generated subtopics
  • an approval flag (yes/no)
• A conditional edge routes execution:
  • If approval == "yes" → proceed to research.
  • If approval == "no" → loop back to subtopic generation.

---

LangChain / LangGraph Implementation Details Taught

1) Basic LangChain step: prompting and invoking an LLM

• Initializes a local model via an LLM interface (mentions using Llama for inference).
• Creates a prompt template.
• Runs the template and retrieves results using a chain with invoke to produce candidate subtopics.

2) Structured outputs to prevent hallucinations and enable parsing

• Uses Pydantic-style schemas (“schemes” referenced in subtitles) to enforce output structure:
  • Subtopic with:
    • title
    • description
  • SubtopicList as a list of Subtopic
• Employs a structured output parser:
  • generates a JSON schema
  • forces the model output to match that schema
• Adds format instructions to the prompt so results are reliably parseable and usable for archive search.

3) LangGraph structure: nodes, state, edges, conditional routing

• Uses a State graph builder with nodes including:
  • Generate subtopics
  • Ask approval (user interaction)
  • Later nodes: Conduct research
  • Later nodes: Create report
• Example edge flow:
  • START → generate_subtopics → ask_approval
  • then conditional routing based on approval (either proceed or loop/regenerate)
• Mentions using LangGraph visualization to inspect graph flow and how loops behave.

---

Research Pipeline Components Added in the “Final” Agent

After subtopics are approved, the graph expands into a fuller research system:

• Document search node

  • Tool: “FCH Archive Papers” (archive-paper search)
  • Inputs: a subtopic string
  • Method: keyword-based archive search
  • Output: top relevant documents (example: “found three documents” per subtopic)
  • Claim: testing showed the search was accurate/good.

• Gap-finding / analysis prompt

  • Instructs the LLM to determine:
    • what the documents show
    • knowledge gaps (what’s missing / what should be researched next)
    • where deeper study is needed

• Report generation node

  • Combines search and analysis results into a final report saved as research_report.md.
  • Report structure (per subtopic) includes:
    • document information
    • identified research gaps
    • summaries and key details of relevant works
    • authors / what was analyzed (as stated)

• Graph state extension

  • The state structure remains similar, but now includes additional research results, along with new nodes and execution steps.

---

Practical Notes from the Run / Demo

• Demo reportedly required a VPN for reliable archive searching (network/internet issue noted).
• Visualization and step-by-step execution confirmed the control flow:
  • generate → approval → loop/regenerate if rejected → proceed when accepted
  • then search/analyze → final report

---

Debugging / Monitoring Tool Mentioned

• SMITH / LangSmith-like tool (“lmith” in subtitles)
  • Used to track and visualize agent logic calls and outputs
  • Shows a web UI after inserting an API token
  • Demonstrates:
    • when subtopics were generated
    • when the model was called
    • outputs such as research gaps and per-step document info

---

Main Speakers / Sources (Inferred from Subtitles)

• Main speaker: tutorial author/narrator (name not provided in subtitles)
• Primary libraries/tools referenced:
  • LangChain
  • LangGraph
  • Pydantic (structured output schemas)
  • Llama (model/inference mentioned)
  • Mistral Instruct 7B (model mentioned)
  • FCH Archive Papers (archive search tool)
  • LangSmith / Smith (monitoring framework/tool mentioned)