Agentic GraphRAG (Vertical Domain)

Design philosophy: a "good-enough" graph, no over-engineering

Microsoft GraphRAG and Neo4j are powerful, but for a single vertical document base they are often a sledgehammer — you stand up a graph database, learn Cypher, maintain a schema. This project goes the other way:

• the graph is a Python dict: {entities: [...], relations: [...]}, no graph database
• retrieval routing is the agent's job: not hard-coded "always graph" or "always vector" — the LLM looks at the question and picks the tool
• extraction carries source grounding: every extraction keeps a char_interval back to the source text, so answers are traceable

Four-stage pipeline

PDF / long doc
  │
  ├─[1] MinerU parse → Markdown
  │
  ├─[2] LangExtract → entities + data metrics + relations (with char_interval)
  │
  ├─[3] dual write
  │       ├─▶ Chroma (vectors)          ← text-embedding-v4, 1024-dim, chunk_size=10
  │       └─▶ knowledge graph (Python dict)  ← {entities, relations}
  │
  └─[4] LangChain create_agent (3 tools)
          vector_search_tool / graph_search_tool / hybrid_search_tool

LangExtract: extraction with char_interval

It uses Google's open-source langextract==1.1.1, with DeepSeek deepseek-chat as the LLM:

import langextract as lx

result = lx.extract(
    text_or_documents=markdown_text,
    prompt_description=prompt,            # extract entities / data metrics / relations
    examples=few_shot_examples,
    model=deepseek_model,                 # api.deepseek.com / deepseek-chat
    fence_output=True,
    use_schema_constraints=False,
    prompt_validation_level=lx.PromptValidationLevel.OFF,
)

Extractions fall into three classes:

Each extraction carries a char_interval (start/end offset) — the basis for answers that cite "which characters of the source" they come from, a hard requirement in medical / legal / compliance work. The course extracts ~21 items on a sample private-loan contract (referencing Civil Code Article 675): 11 entities and 1 relation.

罗密欧与朱丽叶.txt is LangExtract's few-shot tutorial example in this project, not the main corpus — don't be misled by it.

The knowledge graph is just a Python dict

knowledge_graph = {
    "entities": [
        {"name": "lender", "type": "实体", "attributes": {...}},
        {"name": "loan_amount", "type": "数据指标", "value": "..."},
        # ...
    ],
    "relations": [
        {"主体1": "lender", "主体2": "borrower", "关系": "lends_to"},
        # ...
    ],
}

graph_search does no fancy graph algorithms: substring fuzzy match to locate the starting entity, then a 1–2 hop traversal over relations to pull in connected entities. Good enough — and anyone can read it.

A three-tool agent

agent = create_agent(            # LangChain 1.0
    model=ChatOpenAI(..., temperature=0.3),   # DeepSeek
    tools=[vector_search_tool, graph_search_tool, hybrid_search_tool],
)

• vector_search_tool: Chroma semantic search — good for "what does this document say" fact lookups
• graph_search_tool: entity fuzzy match + 1–2 hop relation traversal — good for "how are A and B related / who else is A connected to"
• hybrid_search_tool: runs both and fuses the evidence — good for compound questions needing facts and relations

The point is not "having a graph" — it is that the agent picks the tool from the question. That is the Agentic in Agentic-GraphRAG.

A single agentic query

Q: "How are the lender and borrower related, and how much must the borrower repay?"
  └─> agent_query() runs a ReAct loop
        ├─ LLM judges: compound relation + numeric question → picks hybrid_search_tool
        ├─ vector path: recalls chunks with "loan amount / interest rate"
        ├─ graph path: locates "lender" → hops to "borrower" → hops to "loan amount"
        └─ fuses → answer with char_interval citations
  └─> returns: answer + tool-call evidence (which path, which entities hit)

What this signals

• Judgment on landing GraphRAG: knowing when you need a graph (relations / multi-hop) and when Neo4j is over-engineering
• Agentic retrieval routing: letting the agent pick vector / graph / hybrid per question instead of hard-coding one path
• Source grounding: char_interval makes every claim traceable — an auditability requirement
• Same roots as LangExtractApp, one step further: the extraction + vector layer is shared with the structured-extraction platform; this adds the knowledge graph + multi-hop agent