hana_ai.agents

hana.ai agents provides focused entry points for dataframe operations, SQL operations, and the recommended ContextAgent workflow for conversational forecasting and dataset preparation.

hana_dataframe_agent

hana_dataframe_agent.create_hana_dataframe_agent(llm, df)

Construct a hana-ml agent from an LLM and dataframe.

hana_sql_agent

hana_sql_agent.create_hana_sql_agent(llm, ...)

Create a HANA SQL agent.

hana_sql_agent.create_hana_sql_toolkit(llm, ...)

Create a HANA SQL toolkit.

context_agent

The file-based ContextAgent is the recommended conversational agent for new workflows. It combines Markdown-backed memory, tool calling, runtime skill routing, and command-style memory or skill controls.

context_agent.ContextAgent(llm[, tools, ...])

Context-engineered agent backed by Markdown memory.

Legacy RAG memory

HANAMLRAGAgent and Mem0HANARAGAgent are deprecated. Existing applications must supply non-empty tenant_id and session_id from trusted application identity, not from model output. Reads, cleanup and deletion are scoped to that tenant/session; existing unscoped memory needs deliberate migration.

For HANAMLRAGAgent, vectorstore_path is a FAISS base directory; the agent adds a tenant/session hash. Its default is HANA_AI_DATA_DIR/chat_history_vectorstore or ~/.hana_ai/chat_history_vectorstore. Existing FAISS files require a valid local integrity manifest. Keep indexes and manifests in host-protected storage; hashes do not make untrusted pickle safe. The explicit HANA_ALLOW_DANGEROUS_FAISS_DESERIALIZATION=true override is only for independently verified trusted indexes during migration, not routine operation.

Mem0 ! administration commands default off; enable HANA_MEM0_ALLOW_ADMIN_COMMANDS=true only behind application-enforced admin authentication. This setting does not control ContextAgent's file-memory commands.