Loading Sample ...
Introduction to RAG: Embed and Query PDFs¶
LangDB aims to bridge the gap between data and AI world by running them alongside. You can create an entire LangDB experience just using SQL commands all the way from extraction to creating a chat agent.
This sample demonstrates the following:
- Extract data from PDF using
extract_textcommand. - Create a RAG function that leverages vector search easily using
View - Create a LangDB
modelthat connects toOpenAIAPIs. - Chat with a LangDB model in natural language using
CHATcommand.
Extract semantically chunked data from pdfs and insert into tables.¶
In [1]:
--Create table for storing pdf chunks
CREATE TABLE pdf_chunks (
id UUID DEFAULT generateUUIDv4(),
content `String`
)
ENGINE = MergeTree
ORDER BY (id, content);
In [2]:
INSERT INTO pdf_chunks(content)
select text from chunk(
(
SELECT content FROM
extract_text( path => 'https://d18rn0p25nwr6d.cloudfront.net/CIK-0000320193/faab4555-c69b-438a-aaf7-e09305f87ca3.pdf',
type => 'pdf')
),
chunk_size => 200,
type => 'Word'
)
In [4]:
SELECT count(*) FROM pdf_chunks;
In [5]:
SELECT * FROM pdf_chunks LIMIT 2;
Embeddings¶
LangDB offers a convenient method to generate embeddings using the custom embedding type model function for development and testing purposes.
In [ ]:
CREATE EMBEDDING MODEL generate_embeddings
USING openai(model='text-embedding-ada-002', encoding_format='float')
In [6]:
--Creating a table for storing embeddings
CREATE TABLE pdf_embeddings (
id UUID,
content `String`,
embeddings `Array`(`Float32`)
)
ENGINE = MergeTree
ORDER BY id;
Use generate_embed model to generate embeddings for each chunk and store them into pdf_embeddings table.
In [7]:
INSERT INTO pdf_embeddings
SELECT id, content, embedding FROM generate_embeddings((
SELECT p.id, content
FROM pdf_chunks AS p
LEFT JOIN pdf_embeddings AS pe ON p.id = pe.id
WHERE p.id != pe.id
ORDER BY p.id
), input=>content)
In [8]:
-- check embeddings
SELECT * FROM pdf_embeddings LIMIT 2;
Vector Search¶
View let you create access endpoints that can be used both for API consumption as well as RAG inputs for LLMs.
Here we are creating a view named similar() to perform a vector search against similar chunks to feed into our LLM model.
In [9]:
CREATE VIEW similar(query String "Query to search similar sections in pdf documents") AS
WITH query AS (
SELECT embedding::Array(Float32) AS query FROM generate_embeddings($query)
)
SELECT
p.id,
p.content,
cosineDistance(p.embeddings, query) AS similarity
FROM
pdf_embeddings p
CROSS JOIN
query
ORDER BY
similarity ASC
LIMIT 5
Sample vector search using similar() function:
In [10]:
--returns similar sections in pdf documents matching the input string
SELECT * FROM similar('Apple liabilities') LIMIT 2
In [11]:
SELECT * FROM similar('PLEASE NOTE: THERE IS NO PROOF OF CLAIM FORM FOR') LIMIT 2
Model and Prompt Creation¶
You can dynamically create models and prompts on the fly using CREATE commands that can leverage tools that you have created.
- Here we are attaching
similarfunction as a tool. - You could also nest queries simply using SQL to hard wire the function chaining.
- Prompts can be passed inline or can be created seperately.
Model Creation¶
In [12]:
CREATE MODEL IF NOT EXISTS search(
input
) USING openai(model_name= 'gpt-4o-mini')
PROMPT (system "You are a helpful assistant. You have been provided with similar tool to search over the SEC fillings.
Your task is also to help users with any queries they might have about the SEC filings.
Go through all the content and then respond.",
human "{{input}}")
TOOLS (
similar COMMENT 'View to search embeddings of the PDF'
)
SETTINGS retries = 1;
Model execution¶
Using the created model, we can perform search() query that leverages similar() tool and responds.
This can immediately be plugged into the frontends to also provide a chat interface.
In [ ]:
SELECT * FROM search('Earnings of Apple in 2022')