Retrieval-Augmented Generation (RAG) is the dominant pattern for grounding LLMs in private enterprise data. Most teams reach for Pinecone or Weaviate as the vector store — but if you are already on Oracle, Oracle Database 23ai makes a compelling case to keep everything in one database: ACID transactions, row-level security, SQL joins between your vectors and relational data, and no additional managed service to operate. This post walks through a full production RAG pipeline from raw documents to a working chatbot endpoint.

Architecture Overview

Raw Documents (PDF / HTML / DB rows)
        │
        ▼
  Text Chunker (LangChain RecursiveCharacterTextSplitter)
        │
        ▼
  Embedding Model (OpenAI ada-002 / OCI GenAI embed-english-v3)
        │
        ▼
  Oracle 23ai VECTOR column  ←→  Metadata (source, page, date, dept)
        │
        ▼
  Oracle HNSW Vector Index (COSINE distance)
        │
        ▼
  LangChain OracleVS Retriever
        │
        ▼
  GPT-4o / Llama 3 (context-augmented generation)
        │
        ▼
  FastAPI / Oracle ORDS REST endpoint

1. Database Setup

Create the Knowledge Base Table

-- Oracle 23ai
CREATE TABLE kb_chunks (
  chunk_id     NUMBER GENERATED ALWAYS AS IDENTITY PRIMARY KEY,
  source_doc   VARCHAR2(500)   NOT NULL,   -- filename or URL
  page_num     NUMBER,
  dept         VARCHAR2(100),              -- for row-level filtering
  chunk_text   CLOB            NOT NULL,
  embedding    VECTOR(1536, FLOAT32),      -- OpenAI ada-002 / text-embedding-3-small
  created_at   TIMESTAMP DEFAULT SYSTIMESTAMP
);

-- HNSW index for sub-millisecond nearest-neighbour search
CREATE VECTOR INDEX kb_hnsw_idx
ON kb_chunks (embedding)
ORGANIZATION INMEMORY NEIGHBOR GRAPH
DISTANCE COSINE
WITH TARGET ACCURACY 95
PARAMETERS (type HNSW, neighbors 32, efconstruction 200);

2. Document Ingestion Pipeline (Python)

import oracledb
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain_openai import OpenAIEmbeddings
from langchain_community.vectorstores.oraclevs import OracleVS
from langchain_community.vectorstores.utils import DistanceStrategy
import fitz   # PyMuPDF

# Connect to Oracle 23ai
connection = oracledb.connect(
    user="KNOWLEDGEBASE",
    password="your_password",
    dsn="your_host:1521/FREEPDB1"
)

# Embedding model
embeddings = OpenAIEmbeddings(model="text-embedding-3-small")  # 1536 dims

# Chunk a PDF
def ingest_pdf(filepath: str, dept: str):
    doc = fitz.open(filepath)
    splitter = RecursiveCharacterTextSplitter(
        chunk_size=800,
        chunk_overlap=120,
        separators=["

", "
", ". ", " "]
    )
    for page_num, page in enumerate(doc):
        text = page.get_text()
        chunks = splitter.split_text(text)
        for chunk in chunks:
            # Get embedding vector
            vector = embeddings.embed_query(chunk)
            # Insert into Oracle
            with connection.cursor() as cur:
                cur.execute("""
                    INSERT INTO kb_chunks
                        (source_doc, page_num, dept, chunk_text, embedding)
                    VALUES (:1, :2, :3, :4, to_vector(:5))
                """, [filepath, page_num + 1, dept, chunk,
                      str(vector)])   # Oracle accepts Python list as vector literal
        connection.commit()
    print(f"Ingested {filepath} — {page_num+1} pages")

ingest_pdf("oracle_erp_manual.pdf", "Finance")
ingest_pdf("zatca_compliance_guide.pdf", "Tax")

3. Using LangChain OracleVS (Recommended)

LangChain 0.2+ ships an official OracleVS vector store. It handles table creation, embedding insertion, and similarity search automatically:

from langchain_community.vectorstores.oraclevs import OracleVS
from langchain_community.vectorstores.utils import DistanceStrategy

# Create vector store (creates table if not exists)
vs = OracleVS.from_documents(
    documents=chunks,             # list of LangChain Document objects
    embedding=embeddings,
    client=connection,
    table_name="KB_CHUNKS_LC",
    distance_strategy=DistanceStrategy.COSINE
)

# Similarity search
results = vs.similarity_search(
    "What are the ZATCA Phase 2 XML requirements for B2B invoices?",
    k=5
)
for doc in results:
    print(doc.page_content[:200], "---", doc.metadata)

4. RAG Chain with Conversation Memory

from langchain_openai import ChatOpenAI
from langchain.chains import ConversationalRetrievalChain
from langchain.memory import ConversationBufferWindowMemory

llm = ChatOpenAI(model="gpt-4o", temperature=0.1)

retriever = vs.as_retriever(
    search_type="similarity",
    search_kwargs={"k": 5}
)

memory = ConversationBufferWindowMemory(
    memory_key="chat_history",
    return_messages=True,
    k=6   # keep last 6 turns
)

chain = ConversationalRetrievalChain.from_llm(
    llm=llm,
    retriever=retriever,
    memory=memory,
    verbose=False,
    return_source_documents=True
)

# Ask a question
result = chain.invoke({"question": "Explain Oracle Fusion AP invoice approval workflow"})
print(result["answer"])
for src in result["source_documents"]:
    print(f"  Source: {src.metadata["source_doc"]} p.{src.metadata["page_num"]}")

5. Hybrid Search: Vectors + SQL Filters

Oracle's biggest advantage over standalone vector DBs is the ability to combine vector similarity with SQL predicates in one query — no post-filtering approximations:

-- Retrieve top-5 chunks from Finance dept only, written after Jan 2025
SELECT chunk_id, chunk_text, dept,
       VECTOR_DISTANCE(embedding, to_vector(:query_vec), COSINE) AS distance
FROM   kb_chunks
WHERE  dept = 'Finance'
  AND  created_at > TIMESTAMP '2025-01-01 00:00:00'
ORDER BY distance
FETCH FIRST 5 ROWS ONLY;

This runs the HNSW index scan with the SQL predicate pushed down — Oracle's CBO decides whether to use the vector index or a full scan based on selectivity.

6. Serving as a REST API (Oracle ORDS)

-- Enable REST on the table
BEGIN
  ORDS.ENABLE_OBJECT(
    p_enabled      => TRUE,
    p_schema       => 'KNOWLEDGEBASE',
    p_object       => 'KB_CHUNKS',
    p_object_alias => 'kb'
  );
END;
/

-- Or use ORDS AutoREST + a PL/SQL module for the RAG endpoint
-- POST /ords/knowledgebase/ai/query → calls DBMS_CLOUD_AI internally

Performance Benchmarks

Tested on OCI VM.Standard3.Flex 16 OCPU, 256 GB RAM, Oracle 23ai Free.

Key Takeaways

• Oracle 23ai eliminates the need for a separate vector store for enterprise RAG
• HNSW index delivers sub-5ms similarity search on datasets up to ~5M vectors in memory
• Hybrid SQL + vector queries are Oracle's biggest differentiator over Pinecone / Chroma
• LangChain OracleVS provides a production-ready integration layer
• Row-level security applies automatically to vector search results

ROSTAN Technologies builds RAG systems on Oracle 23ai for enterprise knowledge bases, ERP documentation assistants, and compliance Q&A bots. Book a technical consultation to see a live demo on your Oracle environment.