Text2SQL Release Notes
Solid releases from the last 90 days that changed how Text2SQL turns your questions into SQL, and whether your results could change.
This page lists each Solid release from the last 90 days that changed how Text2SQL turns questions into SQL. Check it when a question you ask regularly starts returning different SQL or results, or when you track answer quality over time and want to know whether a change came from Solid.
What's covered. Text2SQL wherever you use it: through the Solid MCP server (from your own agents, Copilot Studio, Snowflake, and other clients) and in chat inside Solid. Releases that didn't change Text2SQL aren't listed, so if a release date isn't here, Text2SQL behaved the same before and after it.
Matching a query to a release. Every SQL statement Solid generates includes a comment such as Generated by Solid at Oct 26, 2026. The release that produced the query is the most recent entry dated on or before that date. A query generated on the same day as a release may come from either that release or the one before it.
Reading an entry. Each entry shows the release date and version, whether existing questions could now produce different SQL, what changed (New, Improved, Changed, or Fixed), and whether you need to do anything. Entries older than 90 days are removed. If you have questions about a change, contact your Solid team.
Oct 26, 2026 · Version 0.1954.0
Results may change: Yes. Some questions may now produce different SQL or results.
- Fixed: On Databricks, questions that filter on specific values, such as a product name, a status or a numeric range, now use the values actually stored in the column more consistently.
What you might notice: Filters in generated Databricks SQL may now use values written exactly as they appear in your data, and range filters may better reflect the column's actual values.
Action needed: If you track answer quality over time, re-run your benchmark or evaluation questions to see how this release affects them.
Oct 25, 2026 · Version 0.1947.0
Results may change: Yes. Some questions may now produce different SQL or results.
- Changed: Semantic models built or rebuilt from BI reports after this release now include fewer near-duplicate report queries among their example queries, so questions on those models may get different SQL.
What you might notice: Newly built or rebuilt models may list fewer, less repetitive example queries from BI reports. Existing models are unchanged until they are rebuilt.
Action needed: If you track answer quality over time, re-run your benchmark or evaluation questions after rebuilding a semantic model from BI reports.
Oct 22, 2026 · Version 0.1943.0
Results may change: Unlikely. Existing questions should produce the same SQL.
- Fixed: When Solid needs additional tables to answer a question, the answer's explanation, the tables it describes and any note about missing data now match the tables the query actually uses.
What you might notice: Explanations and missing-data notes may now mention tables they previously left out, matching the SQL. The generated SQL itself does not change.
Action needed: None.
Oct 20, 2026 · Version 0.1926.0
Results may change: Yes. Some questions may now produce different SQL or results.
- Improved: Questions on semantic models with golden queries or many custom example queries now draw on those examples more consistently, with golden queries treated as preferred examples.
- Improved: Repeated questions that span more than one semantic model now return the same SQL more consistently.
- Changed: Short, vague or very broad questions may now get a clarifying question where they previously got SQL, or SQL where they previously got a clarifying question.
- Changed: Some questions may now get a different query than in earlier releases, even when the question and semantic model are unchanged.
- Fixed: On SQL Server, questions that filter on text values, especially non-English values, now use the values actually stored in the column more consistently.
What you might notice: Generated SQL may follow your golden and custom example queries more closely, and vague questions may now get SQL or a clarifying question where they previously got the other. On SQL Server, text filters may match stored values in more cases.
Action needed: If you track answer quality over time, re-run your benchmark or evaluation questions to see how this release affects them.
Oct 15, 2026 · Version 0.1898.0
Results may change: Yes. Some questions may now produce different SQL or results.
- Changed: Users with the Viewer role now get chat answers and follow-up suggestions that can draw on tables, queries and metrics, not only dashboards and reports, the same as for Analysts.
- Fixed: Questions on semantic models with columns whose names contain a period can now use those columns in generated SQL; previously those columns were left out.
- Fixed: For columns that your past queries usually filter by excluding certain values, generated SQL is less likely to include those values instead, once query history has been reprocessed.
What you might notice: Columns with a period in their name may now appear in generated SQL, and filters on columns usually filtered by exclusion may now exclude values rather than include them. Viewers may see tables, queries and metrics in chat answers and suggestions.
Action needed: If you track answer quality over time, re-run your benchmark or evaluation questions to see how this release affects them.
Oct 13, 2026 · Version 0.1879.0
Results may change: Yes. Some questions may now produce different SQL or results.
- New: Questions on semantic models built on SAP HANA, which previously failed, now return SQL written for SAP HANA.
- Changed: Questions asked in a language other than English now get replies in that language only when it is enabled for your organization, and in English otherwise.
What you might notice: Replies to non-English questions may now come back in English when that language is not enabled for your organization, and SQL for non-English or mixed-language questions may differ. Questions on SAP HANA models now return SQL instead of an error.
Action needed: If non-English questions now get English replies, ask your Solid contact which languages are enabled for your organization. If you track answer quality over time, re-run your benchmark or evaluation questions to see how this release affects them.
Oct 11, 2026 · Version 0.1870.0
Results may change: Unlikely. Existing questions should produce the same SQL.
- New: The Text2SQL tool on Solid's MCP server now accepts an optional original question field, so your agent can pass the user's exact wording when it has rewritten or expanded the question.
What you might notice: MCP clients will see the new optional field in the tool's definition. Existing calls work as before and the generated SQL does not change.
Action needed: None.
Oct 7, 2026 · Version 0.1860.0
Results may change: Yes. Some questions may now produce different SQL or results.
- Fixed: On SQL Server and Microsoft Fabric Lakehouse, questions that filter on specific values now use the values actually stored in the column more consistently, including for column names with spaces, parentheses or non-Latin characters.
- Fixed: Questions that join on columns whose names contain a period now return the generated query in cases where they previously came back with no SQL.
What you might notice: On SQL Server and Fabric Lakehouse, filters may now use values written exactly as they appear in your data. Questions involving columns with a period in their name may now return SQL where they previously returned none.
Action needed: If you track answer quality over time, re-run your benchmark or evaluation questions to see how this release affects them.
Oct 6, 2026 · Version 0.1854.0
Results may change: Yes. Some questions may now produce different SQL or results.
- New: You can now connect a SAP HANA warehouse as a data source.
- Changed: The 'Generated by Solid at' comment in generated SQL now appears as a block comment right after the first SELECT, instead of on a separate first line; query logic is unchanged.
- Fixed: On Teradata and Starburst (Trino) sources, questions involving INTEGER columns can now take those columns' values and ranges into account when choosing filters.
What you might notice: The text of generated SQL changes slightly because of the comment's new position, which matters if you compare or parse generated SQL over time, though the query logic is unchanged. The same comment appears in your warehouse's query history.
Action needed: If you compare or parse generated SQL text, update checks that expect the comment on the first line. If you track answer quality over time, re-run your benchmark or evaluation questions to see how this release affects them.
Sep 30, 2026 · Version 0.1818.0
Results may change: Yes. Some questions may now produce different SQL or results.
- Fixed: Semantic models built from Power BI, or extended with suggested tables, after this release now set relationship cardinality and treat ID-like columns as dimensions more consistently when tables have multi-column keys.
What you might notice: In models built or extended after this release, relationships between tables with multi-column keys may have a different cardinality, and some ID-like columns may be grouped by as dimensions rather than aggregated as measures, which can change joins and aggregations in generated SQL.
Action needed: If you track answer quality over time, re-run your benchmark or evaluation questions to see how this release affects them.
Sep 27, 2026 · Version 0.1803.0
Results may change: Yes. Some questions may now produce different SQL or results.
- Improved: Generated SQL and clarifying questions now follow the wording of your question more closely.
- Improved: On larger semantic models whose custom instructions say a specific table must be included, generated SQL now follows that instruction more consistently.
- Improved: Semantic models generated or suggested after this release join tables linked on more than one column more consistently, keeping their key columns together in more cases and keeping multiple joins between the same two tables.
- Changed: The written explanation that comes with each answer now describes how the query answers your question, instead of summarizing the results.
- Fixed: Some questions that failed with an error during periods of heavy load now complete instead.
What you might notice: The explanation with each answer now walks through how the query answers your question rather than summarizing results, and generated SQL may follow the wording of your question more closely. Semantic models generated after this release may include different joins for tables linked on several columns.
Action needed: If you track answer quality over time, re-run your benchmark or evaluation questions to see how this release affects them.
Sep 22, 2026 · Version 0.1782.0
Results may change: Yes. Some questions may now produce different SQL or results.
- Fixed: Complex questions, and questions on large semantic models, that could fail with an error instead of returning SQL now return SQL more consistently.
What you might notice: Fewer complex questions end in an error, and some that previously failed now return SQL.
Action needed: If you track answer quality over time, re-run your benchmark or evaluation questions to see how this release affects them.
Sep 16, 2026 · Version 0.1770.0
Results may change: Yes. Some questions may now produce different SQL or results.
- Fixed: Custom instructions that have a condition set now carry that condition with them, so they apply more consistently only to the questions they are meant for.
What you might notice: An instruction with a condition may now shape the SQL only for questions where the condition holds, so SQL for other questions may no longer reflect it.
Action needed: If you track answer quality over time, re-run your benchmark or evaluation questions to see how this release affects them.
Sep 14, 2026 · Version 0.1754.0
Results may change: Yes. Some questions may now produce different SQL or results.
- New: Questions on semantic models backed by Starburst (Trino) now get SQL written in Trino syntax, such as three-part table names, quoted identifiers and Trino-specific functions.
- Improved: Repeat questions that span two semantic models now return faster.
What you might notice: On Starburst-backed models, generated SQL now uses three-part table names and functions that Trino supports. Repeat questions across two semantic models may also get a response sooner, with no change to the SQL.
Action needed: If you track answer quality over time, re-run your benchmark or evaluation questions to see how this release affects them.
Sep 10, 2026 · Version 0.1746.0
Results may change: Yes. Some questions may now produce different SQL or results.
- Improved: Many questions now return SQL faster, and occasionally the query returned for a question may differ from the one returned before.
What you might notice: Responses may arrive sooner. For a small number of questions, the generated SQL may differ from earlier runs even though the question has not changed.
Action needed: If you track answer quality over time, re-run your benchmark or evaluation questions to see how this release affects them.
Sep 3, 2026 · Version 0.1721.0
Results may change: Unlikely. Existing questions should produce the same SQL.
- Improved: Repeat questions on published semantic models that have custom instructions now return faster, especially on large and layered models.
What you might notice: Repeat questions on these models may get a response sooner; the generated SQL itself does not change. Answers that combine several semantic models and tests of unpublished draft models are not affected.
Action needed: None.
Aug 25, 2026 · Version 0.1689.0
Results may change: Yes. Some questions may now produce different SQL or results.
- Fixed: Questions sent with additional instructions are now less likely to get a clarifying question the instructions already answer, and those instructions now also inform which semantic model answers the question.
What you might notice: Requests that include instructions should get SQL more often instead of a clarifying question, and some may be answered from a different semantic model than before.
Action needed: If you track answer quality over time, re-run your benchmark or evaluation questions to see how this release affects them.
Aug 24, 2026 · Version 0.1687.0
Results may change: Yes. Some questions may now produce different SQL or results.
- Improved: When several semantic models are available, Solid now uses the description each model's owner has written to pick the model that answers a question, so domain-specific questions reach the right model more consistently.
What you might notice: Some questions may now be answered from a different semantic model than before, which changes the generated SQL.
Action needed: Make sure each semantic model's description reflects the questions it should answer. If you track answer quality over time, re-run your benchmark or evaluation questions to see how this release affects them.
Aug 17, 2026 · Version 0.1672.0
Results may change: Yes. Some questions may now produce different SQL or results.
- New: Feedback on a Text2SQL answer sent through Solid's MCP server can now include the IDs of the semantic models behind the answer, including several for answers that combine models, so it reaches each model's owners.
- Fixed: Questions that filter on text values, such as a status or region name, in VARCHAR or NVARCHAR columns, and questions on SQL Server DATETIME2 columns, now take the stored values into account more consistently.
What you might notice: On SQL Server, Postgres, MySQL, Teradata and Vertica sources, filters on these columns may now use values written exactly as they appear in your data, once the source has been re-collected.
Action needed: To pick up the fix on an existing source, re-collect it. If you track answer quality over time, re-run your benchmark or evaluation questions to see how this release affects them.
Aug 12, 2026 · Version 0.1658.0
Results may change: Yes. Some questions may now produce different SQL or results.
- New: Text2SQL responses through Solid's MCP server now include a generation ID, and a new feedback tool lets your agent report whether a query was right, optionally with a corrected query.
- New: You can now connect a SQL Server warehouse as a data source; it was previously listed as coming soon.
- Improved: When a question asks for a metric the semantic model lacks, such as a 28-day count when only a 14-day one exists, the result column is now named for the metric actually used.
- Improved: When a generated query fails to run and Solid fixes it, the fixed query now keeps the original metrics, filters, grouping and grain more consistently.
- Changed: Some questions may now get different SQL than in earlier releases, and asking the same question repeatedly may produce more varied SQL from run to run.
What you might notice: When a similar metric stands in for one the semantic model lacks, the notes now explain the substitution, or Solid says the model cannot answer the question. Agents connected through Solid's MCP server may also add a short line after each answer inviting you to flag a wrong query.
Action needed: If you track answer quality over time, re-run your benchmark or evaluation questions to see how this release affects them.
Aug 9, 2026 · Version 0.1644.0
Results may change: Yes. Some questions may now produce different SQL or results.
- Changed: Instructions sent with a Text2SQL request through Solid's MCP server no longer take precedence over the semantic model's custom instructions.
What you might notice: Where instructions sent with a request conflict with a semantic model's custom instructions, generated SQL may now follow the semantic model's instructions instead. Requests sent without instructions are unaffected.
Action needed: If your agent relies on request instructions to override a semantic model's custom instructions, ask the model's owner to update its custom instructions instead. If you track answer quality over time, re-run your benchmark or evaluation questions to see how this release affects them.
Aug 4, 2026 · Version 0.1637.0
Results may change: Yes. Some questions may now produce different SQL or results.
- Fixed: Questions that don't name a semantic model no longer silently use the first available model when Solid cannot match one; Solid now tries again and otherwise asks for a single semantic model ID.
What you might notice: Some questions that previously got SQL from the first available semantic model may now be answered from a different model, or return a message asking you to specify a single semantic model ID.
Action needed: If you see that message, include a single semantic model ID with the question. If you track answer quality over time, re-run your benchmark or evaluation questions to see how this release affects them.
Aug 2, 2026 · Version 0.1630.0
Results may change: Yes. Some questions may now produce different SQL or results.
- New: Text2SQL requests through Solid's MCP server can now include optional instructions, such as required filters, preferred tables, grain, time zone or row limits, which take precedence over the semantic model's custom instructions. The Aug 9, 2026 release removed this precedence.
- Changed: Text2SQL requests through Solid's MCP server with an empty or blank question now return a clear error explaining that a question is required.
What you might notice: When your agent sends instructions, the generated SQL now reflects them within the semantic model's tables and columns, and any it cannot follow are mentioned in the response notes. Requests without instructions are unaffected, though some agents may start passing constraints as instructions instead of in the question.
Action needed: If you track answer quality over time, re-run your benchmark or evaluation questions to see how this release affects them.
Jul 29, 2026 · Version 0.1613.0
Results may change: Yes. Some questions may now produce different SQL or results.
- Improved: Questions on a layered semantic model now follow that model's own custom instructions when they conflict with inherited ones, and answers that combine several semantic models resolve conflicting instructions more consistently.
- Changed: Questions on layered semantic models may now get SQL that uses a different set of tables or metrics than in earlier releases, even when the question and model are unchanged.
What you might notice: On layered semantic models, generated SQL may now reflect the layered model's own instructions where they used to conflict with inherited ones, and may use a different set of tables or metrics than before.
Action needed: If you track answer quality over time, re-run your benchmark or evaluation questions to see how this release affects them.
Updated 10 days ago
