augmanitai

LangGraph AUGMANITAI Example

A simple 3-node LangGraph StateGraph that demonstrates AUGMANITAI term lookup in a structured workflow.

Architecture

The workflow follows a linear pipeline pattern:

1. query_parser — Extracts relevant AUGMANITAI term names from the user’s question using a mock LLM
2. term_retriever — Searches the AUGMANITAI database for full definitions, examples, and domain classification
3. response_composer — Formats the results into a structured, readable response

State Schema

The AugmanitaiState TypedDict tracks:

• query: User’s research question
• query_terms: Extracted term names (output of parser)
• retrieved_terms: Full term records from the database (output of retriever)
• response: Final formatted response (output of composer)

Files

• state.py — State TypedDict definition
• nodes.py — Node function implementations and mock LLM
• stategraph_example.py — Graph construction and invocation
• test_example.py — Pytest tests
• run_example.py — CLI runner for manual testing
• requirements.txt — Dependencies

Running

Install dependencies:

pip install -r requirements.txt

Run example:

python run_example.py

Run tests:

pytest test_example.py -v

Extending

To extend this example:

• Replace the SimpleLLM class in nodes.py with a real LLM (e.g., Anthropic’s Claude)
• Add conditional branching in the graph (e.g., add nodes to handle ambiguous queries)
• Enhance the response composer to format results as JSON, markdown tables, or other formats
• Add error handling nodes for queries that return no results

Key Learning Points

• StateGraph structure: How to define a typed state and linear workflow
• Offline integration: Mock LLM for demonstration without API keys
• Multi-step reasoning: Separating parsing, retrieval, and formatting concerns
• AUGMANITAI API: Using AugmanitaiDB.search() for terminology lookup

---

Bound by the Ethical Disclaimer §1–§20 (DISCLAIMER.md) and the repository licenses (LICENSE). Contact and provider: IMPRESSUM.md.

This site is open source. Improve this page.