Changelog
Version 1.1.26110100
Security / API ChangesPython execution is now disabled by default, including toolkit refreshes and
used_tools='all'. EnablingPythonHanaMLExecToolrequires host-sideallow_dangerous_code=Trueand custom-tool registration. Dataframe agents also require explicit tools by default and no longer insert table rows into the initial prompt. Enabled Python/REPL tools are not sandboxes; use only trusted code and callers.SmartDataFrame and dataframe/query-to-table tools validate a single read-only SELECT before execution. Protected hana-ml SQL paths reject embedded semicolons, including quoted values, because the dependency splits raw SQL text; SELECT inputs may retain one trailing terminator, which is removed before execution. Query-to-table tools still write their destination tables.
SQL agents require
schemaorinclude_tablesand enforce read-only object access across nested queries, CTEs and joins. Unsupported functions, sequences and comma joins are rejected; optimizer hints must end the outer SELECT. These checks apply to SQL-agent queries, not dedicated model-training, storage or deletion tools, which still require appropriate write/execute permissions. See create_hana_sql_agent for supported syntax.Hardened SQL values and identifiers in vector search, retrieval, infrastructure provisioning, memory export, dataset splitting and forecasting. Safe quoting of retrieval schema/procedure names preserves existing HANA naming semantics. Invalid identifiers, unsupported distance/increment options and unsafe values now fail before execution; vector result limits are bounded.
MCP connection administration requires
HANA_ALLOW_RUNTIME_CONN_UPDATE=true; HTTP administration also requires authentication. Non-loopback HTTP servers require a token; SSE is loopback-only without administration. Reload uses the trusted server environment. TLS downgrades require explicit acknowledgement.Environment-built HANA connections default to encryption and certificate validation. MCP HTTP clients disable redirects and reject remote plaintext HTTP unless explicitly overridden. Remote deployments must provide authenticated HTTPS and trusted certificates.
Stdio MCP startup requires trusted executable/argument profiles in
HANA_MCP_STDIO_ALLOWED_PROFILES. Shells, inline code and Windows batch/PowerShell scripts are rejected; approved absolute paths remain supported.RAG storage defaults to
HANA_AI_DATA_DIRor~/.hana_ai. FAISS reload verifies a local integrity manifest; legacy indexes need migration or explicit trust, and incomplete stores are not overwritten. Keep both indexes and manifests in host-protected storage; integrity checks do not make untrusted pickle safe.Memory operations, including cleanup and deletion, require tenant/session scope supplied by the application. Mem0 chat administration defaults off (
HANA_MEM0_ALLOW_ADMIN_COMMANDS). Identity lookup uses structured connection information or the database user, not connection string representations; applications remain responsible for authentication.ContextAgent session directories use bounded, hashed names to prevent traversal and collisions; existing sessions may need migration. Generated artifacts use unique temporary directories and sanitized names: use returned paths and copy files needed after process exit. Certificate/private-key temporary files are cleaned up on error paths.
Automatic raw HTML rendering is disabled in notebooks and reports.
HANA_ALLOW_RAW_HTML_RENDER=trueis only for trusted content; artifact generation remains available, and no HTML sandbox is provided.
Bug FixesGrouped outlier detection now maps date/timestamp keys through integer indexes and database joins, preserving original dates and group values without unsafe value interpolation. This toolkit workaround does not modify the hana-ml dependency.
Grouped AutoML reports failure without saving when no model is produced. Partial success preserves usable models and returns group errors instead of silently reporting complete success.
Installation / MaintenanceUpdated API documentation and examples for security defaults, memory scope and supported tool behavior; added SQL toolkit and MCP client reference entries.
Removed import-time package installation; install required dependencies explicitly, including
hana_ai[nlp]orhana_ai[graph]for optional features. Python 3.10+ is required. Runtime dependencies use compatible ranges; development dependencies and build backend versions are pinned. Lint CI uses pinned action revisions and a hash-checked Python 3.11/Linux lock.Expanded regression coverage for security, SQL, memory, MCP and ContextAgent workflows.
This version includes intentional compatibility restrictions. Upgrade to
hana_ai==1.1.26110100 and redeploy to apply these fixes; configure least-privilege
HANA roles appropriate to each tool and review the applicable migration notes above.
Application checks do not replace database authorization or host isolation.
Version 1.1.26081400
API ChangesSplit
sap-ai-sdk-genout of the default install.pip install hana_aino longer pulls insap-ai-sdk-gen[all]; usepip install hana_ai[all]to opt back into the full generative AI stack.
Version 1.1.26072000
New FunctionsAdded a set of config dict tools to hana_ml_tools, requiring hana-ml 2.29.26072000.
Version 1.1.260714
New FunctionsAdded the
hana_ai.retrievalpackage to host the HANA AI Core retrieval clients used by the MCP tools; the package exposesObjectDiscovery,DataRetrieval, and the sharedRetrievalBaseused by both. The previous internalhana_ai.agents.hana_agentpackage has been removed in favor ofhana_ai.retrieval.Renamed the MCP-exposed retrieval tools to
object_discovery(schema/object metadata retrieval) anddata_retrieval(data retrieval over remote-source-backed procedures), replacing the previousdiscovery_agent/data_agentnames. The corresponding tool classes areObjectDiscoveryToolandDataRetrievalToolinhana_ai.tools.hana_ml_tools.graph_tools.Added ADI / CV^2 based intermittent-demand classification in
TimeSeriesCheck(ts_checkMCP tool). The check now reports ADI and CV^2 alongside the previous statistics and labels each series assmooth,erratic,intermittent, orlumpyusing the standard thresholds (ADI >= 1.32, CV^2 >= 0.49). ContextAgent skills routing was updated accordingly to preferintermittent_forecastwithmethod=constantfor intermittent series andmethod=sporadicfor lumpy series.Added HTTP MCP audit-context plumbing in
hana_ai.tools.toolkit: the toolkit now attaches aMCPAuditMiddlewarefor HTTP transport (viafastmcp.server.middleware) and mirrors the same context capture on the stdio path so both transports emit consistently keyed audit events. IntroducedDEFAULT_MCP_SESSION_CONTEXT_KEYSinhana_ai.tools.hana_ml_tools.utilityas the canonical set of MCP audit fields (client-declared identity, HANA-authenticated identity, tool name, target tables, response size, model storage name/version, timing, and correlation identifiers).Added
ensure_mcp_audit_logandfind_free_porthelpers inhana_ai.tools.hana_ml_tools.utilityto make bootstrapping a local MCP audit-log file and picking a free HTTP port reusable from example scripts.Added
extra_headerson the MCP HTTP client inhana_ai.client.mcp_client. Callers can now pass arbitrary key/value pairs (for example, a routing hint or a trace-id header) that are merged into the outgoinghttpxclient and per-request headers on top of the standard MCP protocol and session headers.
Version 1.1.260701
Bug FixesFixed ContextAgent failure with newer Anthropic Claude deployments (e.g.
anthropic--claude-4.6-sonnet) on SAP Generative AI Hub that rejecttemperatureandtop_pbeing set together; the agent now sanitizes the LLM sampling parameters on construction and dropstop_pwhentemperatureis also set.Fixed ContextAgent hallucinating table, model, and version identifiers (e.g. substituting
SALES_REFUNDS_TESTfor the previously recordedSALES_REFUNEDS_PREDICT) by feedingDECISIONS.mdandTODO.mdback into the per-turnmemory_notesblock alongsideNOTES.mdand hardening the system prompt to forbid inventing identifiers not present in context.Fixed
invalid table name: Could not find table/view PREDICT_RESULT_...errors fromAutomaticTimeSeriesLoadModelAndPredict,AdditiveModelForecastLoadModelAndPredict, andMassiveAutomaticTimeSeriesLoadModelAndPredictwhen the model storage name (or predict input table/schema) contained lower-case characters: user-supplied identifier fragments are now uppercased before being embedded in thePREDICT_RESULT_/REASON_CODE_/PREDICT_ERROR_table names so the created table aligns with HANA's default folding on unquoted references.Fixed ContextAgent hanging notebook cells for many minutes when a tool call got stuck on the HANA/PAL backend (previously
max_iterationsbounded the number of tool calls but not the wall-clock time, so eight stuck iterations could exceed ten minutes). AddedAgentConfig.max_turn_seconds(default 300s, set toNone/0to disable) which runs the executor on a daemon thread with a join timeout; when the deadline fires the current turn short-circuits with an actionable message instead of hanging.
Version 1.1.260410
Bug FixesFixed the line plot tool issue when the column names in the predicted results table are different from the default column names expected by the tool.
Version 1.1.260408
New FunctionsAdded a standalone ContextAgent skills catalog at hana_ai.iagents.skills.md to make skill definitions easier to review, extend, and maintain.
Added markdown-based skill loading in ContextAgent with built-in fallback support for existing deployments.
Added dedicated dataset preparation tools for CSV import and time-ordered train, test, validation table generation in conversational forecasting workflows, including split_table_for_forecasting.
Added grouped additive forecasting tools backed by AdditiveModelForecast(massive=True) for save-and-predict workflows across many related series.
EnhancementsAdded aggregate and file-level ContextAgent memory reset commands for clearing persisted notes, TODOs, decisions, context, chat history, and session summaries without deleting the storage directory.
Expanded ContextAgent skill coverage for dataset preparation, time-series profiling, prediction-result analysis, grouped forecasting, artifact generation, and dataframe-oriented fallback workflows.
Improved fallback skill selection so profiling, post-prediction analysis, grouped forecasting, artifact generation, and SQL comparison requests map more reliably to the intended skills.
Added validation for markdown-defined skills so malformed entries are ignored unless they provide a valid Goal section.
Clarified that dataset splitting currently targets forecasting scenarios and preserves chronology through time-ordered splitting.
Improved forecasting prediction robustness so automatic, additive, and grouped prediction tools proactively validate and repair predict inputs to inference-only columns when users accidentally pass labeled holdout tables.
Improved ContextAgent tool-failure handling with more actionable diagnostics for predict-versus-score input mismatches and backend PAL runtime failures during forecasting workflows.
API ChangesContextAgent now prefers loading skills from the colocated markdown catalog and merges them with the built-in fallback skills.
ContextAgent is the recommended target for future live end-to-end scenario expansion; HANAMLRAGAgent and Mem0HANARAGAgent are no longer recommended for new scenario coverage.
Version 1.0.260331
New FunctionsAdded prediction result analysis skills and dataframe manipulation skills in the context agent to provide better insights and analysis of the predicted results.
Version 1.0.260330
New FunctionsAdded context agent with markdown type memory to support better context management and display in the conversation.
Version 1.0.260319
Bug FixesFixed the hullucination issue of forecast tools in Mem0HANARAGAgent when the input table name is very similar to the prediction result.
Version 1.0.260316
API ChangesRemoved the graph_tools due to the stored procedure missing.
Version 1.0.260212
Bug FixesFixed the version incomatibility issue of langchain and sap-ai-sdk-gen.
Version 1.0.260116
EnhancementsAdded reset_tools method to HANAMLToolkit to reset the toolkit's tools.
Version 1.0.251223
Bug FixesFixed the import issues in hana_ai agents and tools.
Fixed the missing parameter in PALEmbeddings.
Version 1.0.251217
EnhancementsAdded the customized procedure name in CAP generation tool.
Version 1.0.251001
Bug FixesFixed the output parser issue in SmartDataFrame's transform function.
Version 1.0.250930
EnhancementsAdded MassiveAutomaticTimeSeriesFitAndSave, MassiveAutomaticTimeSeriesLoadModelAndPredict, and MassiveAutomaticTimeSeriesLoadModelAndScore tools to support massive time series model training, prediction, and scoring with group_key parameter.
Added MassiveTSOutlierDetection tool to support massive time series outlier detection with group_key parameter.
Added TSMakeFutureTableForMassiveForecastTool to create future tables for massive time series forecasting with group_key parameter.
Added MassiveTimeSeriesCheck tool to perform time series analysis and generate reports for multiple time series with group_key parameter.
Updated HANAMLToolkit to include the new massive time series tools.
API ChangesModified SelectStatementToTableTool to include a 'force' parameter that allows overwriting existing tables.
Changed "Timeseries" to "TimeSeries" in class names for consistency.
Bug FixesFixed an issue for text with special characters in the HANAVectorEmbeddings class.
Version 1.0.250923
EnhancementsEnhanced the outputs of tools when select_statement is too large by creating temporary tables with unique names.
Added additive_model_forecast_tools and intermittent_forecast df tools to the default tools in SmartDataFrame.
Version 1.0.250918
EnhancementsAdded HANA table schema support for tools.
Improved output information for outlier detections.
Removed the hana_connection_context parameter from the HANAMLRAGAgent class and infer it from the tools.
Refine the default value of max_iterations parameter of HANAMLRAGAgent class parameters from 10 to 20.
Change the default value of vector_store_type parameter of HANAMLRAGAgent class from "faiss" to "hanadb".
Change the default value of long_term_db parameter of HANAMLRAGAgent class from sqlite to HANA DB.
Added the embedding_service parameter to the HANAMLRAGAgent class to allow users to pass their own embedding service. The default embedding service has been changed from GenAIHubEmbeddings to HANAVectorEmbeddings.
Added PAL CrossEncoder as the default cross-encoder model for reranking in the HANAMLRAGAgent class. If it is not available, it will fall back to sentence-transformers/all-MiniLM-L6-v2.
Added session_id parameter to the HANAMLRAGAgent class to support multiple sessions in long-term memory. By default, it is set to "global_session".
Removed the restriction to save memory into long term memory when the result is pandas data or large data. Now, all the results will be saved into long term memory with chunking and embeddings.
Deprecated the code template tool and python REPL tool in SmartDataFrame class. Users can use the tools from df_tools as default tools instead.
Version 1.0.250909
New FunctionsAdded TSMakeFutureTableTool to create a future table for time series forecasting.
Added SelectStatementToTableTool to execute a SELECT SQL statement and store the result in a new table.
Version 1.0.250904
Bug FixesFixed the issue of calling code template tool.
Version 1.0.250707
EnhancementsAdded vector_store_type parameter to HANAMLRAGAgent class to support different vector store types, including "hanadb" and "faiss".
Improved the HANAMLRAGAgent class to handle vector store initialization and updates more efficiently.
Bug FixesFixed the parameter issues in HANAMLRAGAgent class by adding rerank_candidates and rerank_k parameters.
Version 1.0.250702
New FunctionsAdded HANAMLRagAgent class to enable Retrieval-Augmented Generation (RAG) capabilities, leveraging a hybrid short-term and long-term memory architecture with CrossEncoder reranking techniques.
Version 1.0.250630
New FunctionsAdded hdi artifacts tool.
Version 1.0.250617
New FunctionsAdded model deletion tool and chat history deletion tool to hanaml Agent.
Version 1.0.250530
EnhancementsAdded unsupported tools check (classfication, regression).
BAS integration enhancements.
Version 1.0.250520
EnhancementsAdded set_return_direct function in hanaml Agent.
Bug FixesFixed the prompt in hanaml Agent to enable tool cal.
Fixed CAP generation temporary location in MacOS.
Version 1.0.250509
EnhancementsOutput the inspect code for BAS integration.
Version 1.0.250506
EnhancementsEnhanced the seasonality detection for additive_model_forecast_tools.
Provide table meta information and supported algorithms in ts_check tool.
Version 1.0.250424
EnhancementsAdded input table check and columns check to avoid stopping the reasoning.
Added samples to ts_check tool.
Bug FixesFixed wrong report filename issue. (hana-ml appends _report.html to the file.)
Version 1.0.250411
EnhancementsSave observations to chat history in HANA ML agent. Added max_observations parameter to control the number of observations saved in the chat history.
Adjust the default value of fetch_data tool to return pandas indirectly to avoid chain stopping due to the tool call of fetch_data in the intermediate step.
Version 1.0.250410
EnhancementsEnhanced the HANA SQL agent to support case-sensitive SQL queries.
Added create_hana_sql_toolkit function to create a toolkit for HANA SQL.
Optimized the chat history management in HANA ML agent.
Bug FixesFixed the accuracy_measure tool issue when evaluation_metric="spec".
Version 1.0.250407
EnhancementsImproved forecast_line_plot tool to automatically detect the confidence if it is not provided.
Serialized the tool's return if it is pandas DataFrame when return_direct is set to False.
Bug FixesFixed the json serialization issue when the tool's return contains Timestamp.
Version 1.0.250403
New FunctionsAdded list_models tool to list all trained models in the model storage.
Added accuracy_measure tool to measure the accuracy of a model on a test dataset for time series forecasting.
EnhancementsImproved the intermittent_forecast tool to use CrostonTSB instead.
Added parameter return_direct to all tools and toolkit.
Improved the fetch_data tool to return a pandas DataFrame instead of a list of dictionaries. By default, the tool parameter return_direct is set to True, which means the tool will return a pandas DataFrame.