Skip to content

Commit

Permalink
WMSID: 11620 - First commit for PLSQL RAG Livelab (#134)
Browse files Browse the repository at this point in the history 
* Update on LAB1 & LAB2

In LAB 1 updated Task 6 & added task 7
Updated LAB2

* Update lab1.md

correcting Task 2 title
  • Loading branch information
@rrumale
rrumale authored May 17, 2024
1 parent f576611 commit 61da99c
Show file tree
Hide file tree
Showing 2 changed files with 194 additions and 14 deletions.
206 changes: 193 additions & 13 deletions plsql-rag-23ai/lab1/lab1.md
View file
Original file line number Diff line number Diff line change
Expand Up @@ -126,7 +126,7 @@ In this lab, you will:
```
## Task 2: Understanding DBMS\_VECTOR\_CHAIN package.
## Task 2: Understanding `DBMS_VECTOR_CHAIN` package.
**1 - Load the document**
Expand Down Expand Up @@ -363,39 +363,39 @@ Lets summarize the steps we did so far.
<tbody>
<tr>
<td>
You have successfuly loaded a pdf file into my\_books.
You have successfuly loaded a pdf file into my_books.
</td>
<td>
DBMS\_LOB.GETLENGTH(to\_blob(bfilename('VEC\_DUMP', 'Release\_notes.pdf')))
DBMS_LOB.GETLENGTH(to_blob(bfilename('VEC_DUMP', 'Release_notes.pdf')))
</td>
</tr>
<tr>
<td>
Then you have opened the pdf and extract the text
</td>
<td>
DBMS\_VECTOR\_CHAIN.UTL\_TO\_TEXT(dt.file\_content))
DBMS_VECTOR_CHAIN.UTL_TO_TEXT(dt.file_content))
</td>
</tr>
<tr>
<td>
next the text is broken up into text\_chunks.
next the text is broken up into text_chunks.
</td>
<td>
DBMS\_VECTOR\_CHAIN.UTL\_TO\_CHUNKS
DBMS_VECTOR_CHAIN.UTL_TO_CHUNKS
</td>
</tr>
<tr>
<td>
next creating vector embeddings for each of the text chunks
</td>
<td>
DBMS\_VECTOR\_CHAIN.UTL\_TO\_EMBEDDING
DBMS_VECTOR_CHAIN.UTL_TO_EMBEDDING
</td>
</tr>
<tr>
<td>
finally inserted text\_chunk, vectors into vector\_store
finally inserted text_chunk, vectors into vector_store
</td>
<td>
INSERT AS SELECT..
Expand Down Expand Up @@ -454,7 +454,11 @@ When presented with text chunk select from the vector search result and a user q
In the code below we are emmbeding the user question, doing a vector seach in vector store for selecting relevant text chunks using vector distance function. we then send the select chunks to Oracle GenAI which provides an repsonse.
For connecting to LLM we have pre-created login credintials using DBMS_VECTOE.CREATE_CREDENTIAL. The syntax of establising connection with Oracle GenAI can be checked in the documentation here.
For connecting to LLM we have pre-created login credintials using DBMS_VECTOE.CREATE_CREDENTIAL. The syntax of establising connection with Oracle GenAI can be checked in the documentation.
The below code is creating a function that does the vector search using VECTOR\_DISTANCE function. And send the resulting text chunks to OCI GenAI LLM service to get answers.
We will use this function in next step to genrate the response.
1. Select the cell and click **Run**.
Expand All @@ -479,7 +483,7 @@ For connecting to LLM we have pre-created login credintials using DBMS_VECTOE.CR
search_query := 'SELECT EMBED_DATA from VECTOR_STORE where doc_id ='||doc_id||' ORDER BY vector_distance(EMBED_VECTOR, :user_question_vec, COSINE)'||
' FETCH FIRST 10 ROWS ONLY ' ;
OPEN message_cursor FOR search_query USING user_question_vec;
--select embed_data from vector_store where doc_id=7;
-- Initialize messages CLOB
messages := '';
Expand Down Expand Up @@ -533,15 +537,191 @@ For connecting to LLM we have pre-created login credintials using DBMS_VECTOE.CR
cursor.execute(procedure_query)
```
2. Select the doc id on which to query
In the function `generate_text_response_gen` we pass the doc_id to restrict the chucks a related to a individual doucment we load. This improves the acquracy of the vector_search of the question.
```
%%sql select a.file_name, a.file_size , b.doc_id,b.count,b.avg_text_chunk_size, b.max_chunk_size from
my_books a, (select doc_id ,count(*) count, round(avg(length(embed_data))) avg_text_chunk_size, max(length(embed_data)) max_chunk_size from vector_store group by doc_id) b
where a.id = b.doc_id
```
3. Generate LLM response
Generate the response using the query below. This call function `generate_text_response_gen` and passed the `doc_id` from previous step
```
%%sql
select generate_text_response_gen('List specific points in the Oracle Database 23c release note', 1) from dual
```
Observer the output. Retry by replacing the question and see how the output change
We successfully performed the vector search, sent the relevant chunks to the LLM along with the question and got the answer.
We things we learned in this lab how to create a complete RAG solution using PLSQL. The tabel below list the Vector funcation we used for operation.
<table>
<thead>
<tr>
<th style={{ width: "200px" }}>
Description
</th>
<th>
Function
</th>
</tr>
</thead>
<tbody>
<tr>
<td>
Emmbeding the user question
</td>
<td>
VECTOR_EMBEDDING(tinybert_model USING user_question as data)
</td>
</tr>
<tr>
<td>
Do vector search on question and pdf chucks embedings
</td>
<td>
VECTOR_DISTANCE(EMBED_VECTOR, :user_question_vec, COSINE)
</td>
</tr>
<tr>
<td>
Pass the result chucks and the user question to the LLM
</td>
<td>
DBMS_VECTOR_CHAIN.UTL_TO_GENERATE_TEXT(messages, json(params_genai))
</td>
</tr>
</tbody>
</table>
**HURRAY!!! we just completed the LAB1**
We saw the complete implementation of RAG using PLSQL.
## Task 7: Preparation for LAB2
In the LAB 2 we will run a python application which uses the funcation and code we just learned in previous steps.
In addition to the above code for the python application we have the follow additional code.
a. Create a procedue `insert_my_table_row` to insert PDF file in MY\_BOOKS table and returns the doc\_id
b. Create a trigger `trg_mybooks_vector_store_compound` to create embedding for the PDF and stores it in VECTOR\_STORE table.
1. Create a procedue `insert_my_table_row`
```
cursor = conn23ai.cursor()
procedure_query = """
create or replace PROCEDURE insert_my_table_row(
p_file_name IN my_books.file_name%TYPE,
p_file_size IN my_books.file_size%TYPE,
p_file_type IN my_books.file_type%TYPE,
p_file_content IN my_books.file_content%TYPE,
p_new_id OUT number
)
IS
v_count NUMBER;
v_id number;
new_id number ;
BEGIN
-- Check if the combination of a and b already exists
BEGIN
SELECT id INTO new_id FROM MY_BOOKS WHERE file_name = p_file_name AND file_size = p_file_size;
EXCEPTION WHEN NO_DATA_FOUND THEN
INSERT INTO MY_BOOKS (file_name, file_size, file_type, file_content)
VALUES (p_file_name, p_file_size, p_file_type, p_file_content)
RETURNING id into new_id;
END;
p_new_id := new_id;
dbms_output.put_line(new_id);
COMMIT;
EXCEPTION
WHEN OTHERS THEN
-- Exception handling here, for example, a rollback or a custom error message
DBMS_OUTPUT.PUT_LINE('Error: ' || SQLERRM);
--RAISE;
END insert_my_table_row;
"""
# Execute the procedure creation query
cursor.execute(procedure_query)
```
2. Create a trigger `trg_mybooks_vector_store_compound`
```
cursor = conn23ai.cursor()
create_trigger = """
CREATE OR REPLACE TRIGGER trg_mybooks_vector_store_compound
FOR INSERT ON my_books
COMPOUND TRIGGER
TYPE t_id_tab IS TABLE OF my_books.id%TYPE INDEX BY PLS_INTEGER;
v_ids t_id_tab;
AFTER EACH ROW IS
BEGIN
v_ids(v_ids.COUNT + 1) := :NEW.id;
END AFTER EACH ROW;
AFTER STATEMENT IS
BEGIN
FOR i IN 1 .. v_ids.COUNT LOOP
INSERT INTO vector_store (doc_id, embed_id, embed_data, embed_vector)
SELECT dt.id AS doc_id,
et.embed_id,
et.embed_data,
to_vector(et.embed_vector) AS embed_vector
FROM my_books dt
CROSS JOIN TABLE(
dbms_vector_chain.utl_to_embeddings(
dbms_vector_chain.utl_to_chunks(
dbms_vector_chain.utl_to_text(dt.file_content),
json('{"by":"words","max":"300","split":"sentence","normalize":"all"}')
),
json('{"provider":"database", "model":"tinybert_model"}')
)
) t
CROSS JOIN JSON_TABLE(
t.column_value,
'$[*]' COLUMNS (
embed_id NUMBER PATH '$.embed_id',
embed_data VARCHAR2(4000) PATH '$.embed_data',
embed_vector CLOB PATH '$.embed_vector'
)
) AS et
WHERE dt.id = v_ids(i);
END LOOP;
END AFTER STATEMENT;
END trg_mybooks_vector_store_compound;
"""
# Execute the trigger creation
cursor.execute(create_trigger)
```
## Summary
Run the both the steps and proceed to lab 2
You may now [proceed to the next lab](#next).
## Acknowledgements
* **Authors** - Vijay Balebail, Milton Wan, Douglas Hood, Rajeev Rumale
* **Last Updated By/Date** - Milton Wan, May 2024
* **Authors** - Vijay Balebail, Milton Wan, Rajeev Rumale
* **Last Updated By/Date** - Vijay Balebail, May 2024
2 changes: 1 addition & 1 deletion plsql-rag-23ai/lab2/lab2.md
View file
Original file line number Diff line number Diff line change
Expand Up @@ -80,4 +80,4 @@ By using AI Vector Search in Oracle Database 23ai, you can build RAG application
## Acknowledgements
* **Authors** - Vijay Balebail, Milton Wan, Rajeev Rumale
* **Last Updated By/Date** - Rajeev Rumale, May 2024
* **Last Updated By/Date** - Milton Wan, May 2024

0 comments on commit 61da99c

Please sign in to comment.