hi i tried the example shown here

also related question, why is "Summary" needed for structured data like a SQL index? makes sense why it is needed for chunkifying unstructured data like a notion doc

to be clear, i'm just trying to use llama-index for natural language to SQL. i started trying the GPTTreeIndex and the GPTListIndex to run join queries across base indices. should i be taking a different approach? @jerryjliu0

hi, @alex you want to use the SQL index right? for the SQL index you shouldn't be customizing the mode, just leave it as "default" - that's our default text-to-SQL interface

thanks for the quick reply! is there a way to increase the completion token limit? SQL generated from the summary prompt is being cut off in the middle because my table has too many columns

yep! https://gpt-index.readthedocs.io/en/latest/how_to/custom_llms.html#example-changing-the-number-of-output-tokens-for-openai-cohere-ai21

thanks!

also, what are the summaries from these GPTSQLStructStoreIndex objects used for? i see that its generating a bunch of stuff like sums and averages. doesnt it make more sense to just get a description for each table, and feed that to a top level node or something so it has context to perform join queries?

@alex to clarify, what do you mean by summaries? (also i might head off to bed soon but will respond tmrw)

the input for BaseGPTIndex.set_text() , which needs to be set when creating combined indices

Ah I see. You can set the summaries to whatever you want!

We need summaries primarily for composability - since that allows you to define indices on top of other indices, each subindex needs to correspond to some "text" so that the top-level index can index over that

@jerryjliu0 ooh i think i know how to use the SQL Index now...i tried following along with this screenshot in your docs/twitter thread but was really confused about how the GPTStructStoreIndex only accepts a single table_name value, even if the index is being used for JOIN queries

i just picked an arbitrary table in my db, built the context and saved it on the index with store_context_str=True, and queried that 1 SQL Index obj and it seems to be doing pretty well so far

any tips on how to deal with hallucinations? i notice it'll sometimes make up a column or table that doesnt exist

Hi @alex yeah sorry it's really confusing, I'll look into cleaning this up. The table_name field on the index constructor is only needed if you're looking to infer structured datapoints from unstructured data. For text-to-SQL every table included in the SQLDatabase will be used in the prompt.

@jerryjliu0 is there a way to customize the prompt template for PTStructStoreIndex? im currently hacking around it by hardcoded extra context into my query string itself

yeah you can pass in a custom text-to-sql prompt: https://gpt-index.readthedocs.io/en/latest/reference/prompts.html#gpt_index.prompts.prompts.TextToSQLPrompt

pass it in as part of the index.query call: https://gpt-index.readthedocs.io/en/latest/reference/indices/struct_store_query.html

im using BaseGPTIndex.query (using a GPTSQLStructStoreIndex, not GPTNLStructStoreIndexQuery ). looks like its only for the latter @jerryjliu0

yeah. the former is literally just a raw SQL query (it's actually not that useful since you can also just query the SQL engine directly)

sorry if this is stupid question, i am completely new to python, but these AI stuff are mindblowing, so i just cant resist to try something.
so, my question is: how should I pass that text_to_sql_prompt to index.query call?

index.query("bla-bla-bla", text_to_sql_prompt=PROMPT) - this leads to error that text_to_sql is not allowed in query call 😦

hi @TopClans (Азат) , 1) are you using the sql index, and 2) what prompt are you using? let me know if you have a code snippet

1) im using GPTNLStructStoreIndexQuery (hope i answered right)
2) PROMPT_TMPL = ( "This is mysql database schema. \n" "---------------------\n" "{schema}" "\n---------------------\n" "Given this information, please return SQL query, and nothing else: {query_str}\n" ) PROMPT = TextToSQLPrompt(PROMPT_TMPL)

btw, there is an example: https://gpt-index.readthedocs.io/en/stable/reference/indices/struct_store_query.html#gpt_index.indices.query.struct_store.GPTNLStructStoreIndexQuery

response = index.query("<query_str>", mode="sql")

when i add mode="sql", it says:
TypeError: query() got an unexpected keyword argument 'mode'

hi @TopClans (Азат) , what is your index variable? So, GPTNLStructStoreIndexQuery is not public facing, you would instead do something like this:

Setting mode="default" will use the GPTNLStructStoreIndexQuery class under the hood

this is my code:
PROMPT = TextToSQLPrompt(PROMPT_TMPL) prompt_helper = PromptHelper(4096, 4096, 10) index = GPTNLStructStoreIndexQuery( [], sql_database=sql_database, sql_context_container=context_container, prompt_helper=prompt_helper ) response = index.query(query_str, mode="default", prompt=PROMPT)

using index via GPTSQLStructStoreIndex with param prompt:
index = GPTSQLStructStoreIndex( [], sql_database=sql_database, table_name="1c_sales_items", )
response = index.query("blablabla?", mode="default", prompt=PROMPT)

and getting error:
TypeError: __init__() got an unexpected keyword argument 'prompt'

and if dont use that prompt, generated SQL query consist ambiguous field

btw I watched your podcast with Weaviate yesterday and you described some mind-blowing opportunities for devs, thats very impressive technology

sorry i named the prompt field wrong

try index.query(..., text_to_sql_prompt=...)