Requesting a structured output

When creating lexical graphs, you extracted the character name and the lines from the text using a text splitter, but this approach often results in lost context around the split boundaries. For example, the stage directions contained in the text can be used to determine who the character is speaking to at the time.

You'll use Character class inside a new Line class to describe the line being spoken, along with the speaker and a list of characters the line is being spoken to. You'll also create a wrapper LineOutput class to extract multiple of these lines at once.

The prompt variable containing extraction instructions has been defined for you, along with an instance of ChatOpenAI as llm.

This exercise is part of the course

Graph RAG with LangChain and Neo4j

Exercise instructions

Update the LineOutput class to request a list of Line objects.
Define an LLM that generates structured outputs using the LineOutput class.

Hands-on interactive exercise

Have a go at this exercise by completing this sample code.

class Line(BaseModel):
    text: str = Field(..., description="The text of the line")
    spoken_by: Character = Field(..., description="The character who speaks the line")
    spoken_to: Optional[list[Character]] = Field(None, description="The character who the line is spoken to")

class LineOutput(BaseModel):
    # Request more than one line from the text
    lines: ____[Line] = Field(..., description="The lines from the text")

# Return an instance of LineOutput
structured_llm = llm.____(____)

for line in structured_llm.invoke(prompt.format_messages(text=text)):
	print(line)

Edit and Run Code

This exercise is part of the course

Graph RAG with LangChain and Neo4j

AdvancedSkill Level

4.9+

Start Course for Free

Learn how Graph RAG can improve the accuracy and reliability of RAG applications! Store information as nodes and edges in a Neo4j database, and give your LLM the ability to query it so it can retrieve entity and relational information to provide informed answers.

Exercise 1: Graphs and RAG Exercise 2: Creating nodes Exercise 3: Creating relationships Exercise 4: Saving graph documents Exercise 5: Querying a knowledge graph Exercise 6: Writing Cypher statements Exercise 7: Running Cypher statements Exercise 8: The MERGE clause Exercise 9: Text-to-Cypher Graph RAG with Neo4j Exercise 10: Building a text-to-Cypher chain Exercise 11: Text-to-Cypher retrieval chain

Text-to-Cypher applications work well in many circumstances, but we can do better than that! Discover how to construct graph databases using different graph models including lexical and domain graphs. Create Neo4j vector indexes so that you can have the best of both worlds and run graph and vector retrieval simultaneously!

Exercise 1: Lexical graphs Exercise 2: Elements of a lexical graph Exercise 3: Splitting the play into Acts Exercise 4: Creating a hierarchical lexical graph Exercise 5: Combining lexical graphs with vector search Exercise 6: Creating text chunks Exercise 7: Creating a vector index Exercise 8: Domain graphs Exercise 9: Creating a structured output Exercise 10: Requesting a structured output

Current Exercise

Exercise 11: Providing few-shot examples Exercise 12: Building a hybrid retrieval chain Exercise 13: Runnable lambdas Exercise 14: Assigning additional values to an input Exercise 15: The final link in the chain

Although Graph RAG applications are generally more reliable than vector RAG, they aren't totally infallible. In this chapter, you'll learn to evaluate your Graph RAG applications, spot incorrect or duplicate graph nodes, and integrate long-term memory so user preferences can be learned over time.

Exercise 1: Entity resolution Exercise 2: Using extracted entities Exercise 3: Graph-based entity resolution Exercise 4: Evaluating Graph RAG with RAGAS Exercise 5: Creating a Ragas evaluation Exercise 6: Evaluating context retrieval with Ragas Exercise 7: Memory graphs Exercise 8: Saving conversation memory in the graph Exercise 9: Extracting facts from conversation histories Exercise 10: Using extracted conversation facts Exercise 11: Congratulations!