docs/tutorials/sql_qa/

[giscus[bot]]

docs/tutorials/sql_qa/

This guide assumes familiarity with the following concepts:

https://python.langchain.com/docs/tutorials/sql_qa/

[ksteward]

I noticed that this documentation / tutorial listed a bunch of LLMs, but did not list any Llama-3* models.
Do we know if LangChain's SQL database module(s) will work with any Llama-3* models? My initial tests suggested "no" but could well be wrong. I can successfully uses the modules with OpenAI GPT-4o.

[Hamza5]

The official tutorials always use OpenAI.
Llama-3 can be used through Groq. I used the 8b model while following the tutorial and it worked with some problems. It was not intelligent enough to construct the correct SQL that matches my description. It always hallucinated some table and column names that didn't exist. Even though agent corrects itself after getting an error, it seems that it fails to understand what I want.

[llkoolbear]

I was able to successfully use Ollama with qwen2.5:32b and qwen2.5-coder:32b
Was unable to do so with llama3.2
Also the non tool enabled models like deepseek or phi4 couldn't do it.

[ianzhaoyue]

I stuck at the last step of convert-question-to-sql-query, where it gives me an ValueError: no signature found for builtin type <class 'dict'> when invoking write_query({"question": "How many Employees are there?"})

[SharoMonk]

I have particularly found out that various LLMs have varying capacities and capabilities. First thing i do anytime i encountered an error is to change my model. It solves my problems most of the times.
I uses ChatGroq, ChatFireworks, ChatMistrai and ChatXAI models in that order and i have recently added ChatGoogleGenerativeAI. Try it.

[carlosmontoliu]

Excellent! Working smoothly inside a complex Azure WebApp :D

[ksteward]

@carlosmontoliu, are you saying that you CAN get a Llama-3 model to successfully interact with a SQL database? If so, can you reveal which LLM model and which langchain module you used?
Thanks!
-k

[SharoMonk]

I used ChatGroq(model='llama3-8b-8192') and it worked with very little glitches.

[Akinola92]

Can anyone help me figure out why the following isn't retuning any table names after I've saved the db to my Google Drive?

print(db.get_usable_table_names())

Thanks

[Akinola92]

Figured out it was due to the encoding of the Chinook_Sqlite.sql file !

[ksteward]

SharoMonk, So are you saying that model='llama3-8b-8192' successfully used the Langchain SQL database module? from my cell phone -- please excuse typos

…

On Mon, Dec 30, 2024, 11:44 AM SharoMonk ***@***.***> wrote: I used ChatGroq(model='llama3-8b-8192') and it worked with very little glitches. — Reply to this email directly, view it on GitHub <#28698 (reply in thread)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AABOAVSBSVGRTVWKTFGVTWT2IFZ6PAVCNFSM6AAAAABTQ4HDT6VHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTCNRZHAZTINA> . You are receiving this because you commented.Message ID: ***@***.*** com>

[punsoca]

Hey guys- does this section replace the SQLDatabaseChain? I had been trying out the SQLDatabaseChain YouTube tutorials from as recent as August 2024 but my output no longer matching the tutorials.

[ccurme]

This approach is recommended over SQLDatabaseChain. That class parses SQL queries out of model responses. In that sense it is compatible with legacy text-in, text-out LLMs, but does not take advantage of tool-calling and other features of modern chat models. It also does not support the streaming of individual steps or the human-in-the-loop features that you get with the LangGraph implementation. For these reasons SQLDatabaseChain is not featured in our core python documentation.

[sreenathelloti]

How to integrate dynamicfewhot prompting while implementing multi agetn supervisor workflow

[erodriguezds]

Does any of the "Agents" examples work for anybody? I'm using OpenAI LLM, I have followed all of the examples from the "Agents" section, without modification. It blows up inside the stream iteration. I get to see the first step (output from step["messages"][-1].pretty_print()), and then it blows up with the following error:
openai.BadRequestError: Error code: 400 - {'error': {'message': "Invalid type for 'messages[1].content[0]': expected an object, but got a string instead.", 'type': 'invalid_request_error', 'param': 'messages[1].content[0]', 'code': 'invalid_type'}}

[ccurme]

Feel free to share code that is generating the error.

If you are using the OpenAI LLM in this sense:

from langchain_openai import OpenAI

You should use the chat model instead as is done in the doc. The LLM class is a legacy class from when LLMs were primarily text-in, text-out.

[erodriguezds]

Hi @ccurme , I've ended up setting up a public repo to share the code, using the chinook sample database, and it worked smoothly. Then I switched the db connection string to use my database, and it worked too!!! Next, I copied the working code from my test repo to my broken repo, and it worked too!!! So... clearly, I messed up. I'm so sorry for the confusion.

For the record, I'd like to share that I got all the use cases from the Agents page working, including the Dealing with high-cardinality columns. Anyone can take a look a the requirements.txt file to see the packages versions.

[YanzhongSu]

In system message, it says Only use the below tools. which were not listed inside the same prompt. My guess the function create_react_agent will attach such information to the system message based on the specific tools passed in when initialising agent. Is this correct?

agent_executor = create_react_agent(llm, tools, state_modifier=system_message)

[pengfan7758258]

I also want to know about this

[DanialZohairen]

I'm trying to create the agent following the guide (using ChatVertexAI as my LLM) but keep on getting this same error when executing create_react_agent(...) -> TypeError: GenerateSchema.__init__() got an unexpected keyword argument 'types_namespace'

[llkoolbear]

If using FAISS for your vector store, you'll need to pass index and docstore type which can be determined by vectorizing a sample query with your chosen embeddings. Example code below:

import faiss
from langchain_community.vectorstores import FAISS
from langchain_community.docstore.in_memory import InMemoryDocstore
from langchain.agents.agent_toolkits import create_retriever_tool

# Vectorize a sample query
vector_1 = embeddings.embed_query(artists[0])
vector_2 = embeddings.embed_query(albums[0])

# assert and inspect the the vector length
assert len(vector_1) == len(vector_2)
dimension = len(vector_1)
print(f"Generated vectors of length {dimension}\n")
print(vector_1[:10])

# Initialize vector store
index = faiss.IndexFlatL2(dimension)
vector_store = FAISS(
    embedding_function=embeddings,
    index=index,
    docstore=InMemoryDocstore(),
    index_to_docstore_id={},
)

# Add the texts to the vector store and initialize the retriever
_ = vector_store.add_texts(artists + albums)
retriever = vector_store.as_retriever(search_kwargs={"k": 5})
description = (
    "Use to look up values to filter on. Input is an approximate spelling "
    "of the proper noun, output is valid proper nouns. Use the noun most "
    "similar to the search."
)
retriever_tool = create_retriever_tool(
    retriever,
    name="search_proper_nouns",
    description=description,
)

[pengfan7758258]

I did it this way, as follows:

vector_store = FAISS.from_texts(artists + albums, embedding=embeddings)

[praveen-prog]

This code is working fine and I have developed a sample chatbot.
https://github.com/praveen-prog/reltiosqlchatbotv4

[highbrow-228]

Hey,
there is some mistakes with agents in this workflow, because I agent don't see previous results

There is some example:

================================ Human Message =================================

How many records on customers
None
================================== Ai Message ==================================

To find the number of records in the `customers` table, you can use the following SQL query:

```sql
SELECT COUNT(*) AS customer_count FROM customers;

This query will return the total number of rows in the customers table.
None

Please fix it.

[highbrow-228]

bad view of output
here is correct

================================ Human Message =================================

How many records on customers
None
================================== Ai Message ==================================

To find the number of records in the `customers` table, you can use the following SQL query:

"""sql
SELECT COUNT(*) AS customer_count FROM customers;
"""

This query will return the total number of rows in the `customers` table.
None

[AlexSandoval26]

I'm creating an agent which connects to Databricks database and it has some tables with json columns, at time to create a SQL query on that columns the agent don't create correctly the query. What is the correct way to declare this type of queries?

[Rohit10jr]

great tutorial, thank you.

[emalaga]

This documentation is great. For those interested in learning about SQL Agents with LangChain I have created this Tutorial, showing how to build one and analyzing how the agents behaves in a set of increasingly difficult scenarios.