LANGCHAIN — How to Construct a Query

Talk is cheap. Show me the code. — Linus Torvalds

When constructing queries for natural language understanding interfaces (LUI), different data types present specific challenges. Structured, semi-structured, and unstructured data all require tailored query construction processes. In this tutorial, we will explore different approaches for query construction using code snippets and examples.

Text-to-metadata-filter

Vectorstores equipped with metadata filtering enable structured queries to filter embedded unstructured documents. The self-query retriever can translate natural language queries into structured queries using a few steps:

prompt = get_query_constructor_prompt(document_content_description, metadata_field_info)
output_parser = StructuredQueryOutputParser.from_components()
query_constructor = prompt | llm | output_parser

query_constructor.invoke({
 "query": "Songs by Taylor Swift or Katy Perry about teenage romance under 3 minutes long in the dance pop genre"
})

Text-to-SQL

Translating natural language into SQL requests requires addressing challenges such as hallucination and user errors. To ground SQL queries, an LLM can be provided with an accurate description of the database and few-shot examples of question-query matches to improve query generation accuracy. Error handling tools, like SQL agents, can help recover from errors.

Text-To-SQL+semantic

The addition of vector support to relational databases enables semantic searches using the pgvector extension for PostgreSQL. This allows for similarity search over embeddings vector columns and enhances text-to-SQL with knowledge of the semantic operator.

SELECT * FROM tracks ORDER BY "name_embedding" <-> {sadness_embedding}

Text-to-Cypher

Graph databases often use Cypher query language to model relationships between data. Text-to-Cypher can translate natural language to Cypher queries using performant LLMs like GPT-4 to generate Cypher statements.

cypher_chain = GraphCypherQAChain.from_llm(
    cypher_llm = ChatOpenAI(temperature=0, model_name='gpt-4'),
    qa_llm = ChatOpenAI(temperature=0), graph=graph, verbose=True,
)
cypher_chain.run(
    "How many open tickets there are?"
)

Seamlessly retrieving structured and unstructured data across a variety of data sources is crucial for unlocking the potential of LLMs. The provided code snippets and examples offer a starting point for users to construct queries tailored to their specific data types.