oracle-ai-developer-hub
LangChain Ecosystem
A collection of notebooks that build AI applications and agents with the LangChain ecosystem — LangChain, LangGraph, Deep Agents, and Oracle’s first-party integrations (langchain-oracledb, langgraph-oracledb, langchain-oci) — using Oracle AI Database as the single backend for vectors, agent memory, checkpoints, the LLM cache, and chat history.
The theme of the folder is convergence: instead of stitching a vector database, a key-value store, a checkpoint file, and a chat log into one agent, every notebook here keeps all of that state in one converged database.
The stack
Every notebook follows the same shape:
| Layer | Typical component | Role |
|---|---|---|
| Python | the notebook runtime | the glue |
| LLM provider | ChatAnthropic (Claude) or an OpenAI chat model |
the reasoning engine that plans, calls tools, and synthesizes answers |
| LangChain / LangGraph | langchain, langgraph, deepagents, langchain-oracledb, langgraph-oracledb, langchain-oci |
the orchestration fabric and the Oracle-backed persistence primitives |
| Oracle | Oracle AI Database 23ai / 26ai | one converged store for vectors, long-term memory, checkpoints, the LLM cache, and chat history |
Embeddings are a pluggable vectorizer — OpenAI text-embedding-3-small or a local HuggingFace model, depending on the notebook. Each notebook names its own model choices.
The Oracle × LangChain primitives these notebooks use
| Primitive | Package | What it does |
|---|---|---|
OracleVS |
langchain-oracledb |
LangChain vector store backed by Oracle AI Vector Search |
OracleEmbeddings |
langchain-oracledb |
generate embeddings inside the database from an ONNX model |
create_text_index |
langchain-oracledb |
Oracle Text keyword index for hybrid (vector + lexical) search |
OracleSemanticCache |
langchain-oracledb |
reuse an LLM answer when a new prompt is semantically equivalent |
OracleChatMessageHistory |
langchain-oracledb |
durable, session_id-scoped chat transcripts |
OracleStore / AsyncOracleStore |
langgraph-oracledb |
long-term, cross-thread agent memory |
OracleSaver / AsyncOracleSaver |
langgraph-oracledb |
per-thread short-term checkpoints (resume a conversation) |
create_deepagents_agent |
langchain-oci |
Deep Agents factory that auto-wires Oracle datastores into agent tools |
Notebooks
| # | Name | Description | Stack | Open Notebook | Open in Colab |
|---|---|---|---|---|---|
| 01 | RAG with Oracle & LangChain | The starter: build a minimal RAG application on Oracle AI Vector Search with OracleVS and LangChain — ingest documents, embed them, and answer questions grounded in the retrieved chunks. |
Oracle AI Database · langchain-oracledb · HuggingFace |  |
 |
| 02 | Semantic Cache + Chat History | Two langchain-oracledb primitives for cheaper, coherent apps: OracleSemanticCache reuses an answer when a new question means the same as a past one, and OracleChatMessageHistory persists durable, session-scoped transcripts. Includes calibrating the cache’s distance threshold to the embedding model. |
langchain-oracledb · OpenAI (embeddings) · Anthropic | |
|
| 03 | Research Agent with Deep Agents | A durable research agent built with the released langchain_oci.create_deepagents_agent factory. Oracle backs the hybrid-search corpus (ADB), per-thread checkpoints (OracleSaver), long-term notes (OracleStore), and the agent’s virtual filesystem (StoreBackend) — while Claude reasons and plans. |
langchain-oci (deepagents) · Anthropic · HuggingFace · langgraph-oracledb · Oracle AI Database | |
|
| 04 | AI On-Call Triage Assistant | The capstone: a LangGraph supervisor orchestrates two specialists (issue_analyst, policy_agent) to triage incoming issues. Vector knowledge (OracleVS), long-term memory (AsyncOracleStore), checkpoints (AsyncOracleSaver), a semantic LLM cache (OracleSemanticCache), and chat history all share one Oracle AI Database. Runs on Claude Opus 4.8 with adaptive thinking. |
langchain · langgraph · langgraph-supervisor · langchain-oracledb · langgraph-oracledb · Anthropic (Claude Opus 4.8) · OpenAI | |
|
Recommended reading order
The notebooks build on each other from a starter app to a full multi-agent system:
- RAG with Oracle & LangChain (01) — learn the base pattern:
OracleVS, embeddings, retrieval. - Semantic Cache + Chat History (02) — add the two primitives that make an app cheap to run and coherent over time.
- Research Agent with Deep Agents (03) — graduate to an autonomous, planning agent whose memory lives in Oracle.
- AI On-Call Triage Assistant (04) — put it all together: a multi-agent supervisor over five kinds of Oracle-backed state.
Prerequisites
- An Oracle AI Database instance (23ai / 26ai) — either local (each notebook pins a
gvenzl/oracle-freeDocker image and includes thedocker runcommand) or reachable over the network. - Python 3.11+ and Jupyter / JupyterLab (or open any notebook directly in Colab via the badge above).
- API keys, depending on the notebook: an Anthropic key for Claude, and/or an OpenAI key for
text-embedding-3-smallembeddings. Notebook 01 runs entirely on a local HuggingFace embedding model and needs no API key. - Each notebook starts with a single
pip installof the latest packages and lists exactly what it needs.