docs/docs/changelog/Released_V0.8.2.md
In complex data-analysis tasks, business information is often distributed across multiple files and knowledge sources. The analysis may also include work that can proceed in parallel and critical conditions that require confirmation from business users. When file handling, knowledge retrieval, task execution, and human confirmation are disconnected, the Agent struggles to preserve context and the workflow is repeatedly interrupted.
V0.8.2 addresses this end-to-end workflow from input and retrieval through execution and confirmation. Multi-file Agentic analysis manages related inputs together, Agentic Knowledge-Base Search keeps looking for relevant information as the task develops, parallel sub-agent delegation advances independent work at the same time, and human-in-the-loop questions collect decisions that only the user can provide.
Data work often starts with a group of related files rather than one isolated spreadsheet: an orders table and a customer table, several monthly exports, or a report accompanied by supporting data. V0.8.2 lets users attach multiple files to one conversation and keeps that set available as part of the task context.
The default extension set covers CSV, TSV, Excel, JSON/JSONL, Parquet, PDF, Word, PowerPoint, Markdown, and plain text files. Availability of a preview parser still depends on the installed optional dependencies.
The Agent receives stable file identifiers rather than client-provided server paths. Files are scoped to the owning user and the current conversation, then materialized only when an execution tool needs them. This allows load_file, Code Interpreter, and analysis tools to work across the selected file set while keeping storage locations private.
When a conversation with attachments is saved as a Scheduled Task, DB-GPT freezes task-scoped copies of those files. Each scheduled run can therefore replay against the same file snapshot instead of depending on mutable session uploads.
Traditional RAG usually retrieves once, builds a prompt, and generates an answer. That approach is efficient for simple questions, but it gives the model little room to recover when the first query is incomplete or the relevant evidence is spread across several sources.
V0.8.2 turns knowledge-base chat into an Agentic loop. The Agent can inspect the knowledge space, rewrite or narrow its search, retrieve more than once, open relevant files, and stop when it has enough evidence to answer. Knowledge-only chat receives a focused tool set such as semantic search, file listing, glob, grep, and file reading, reducing interference from unrelated tools.
| Capability | How it is used |
|---|---|
| Vector search | Semantic similarity over embedded chunks |
| File and exact search | File matching, keyword search, and bounded file reading within the selected knowledge space |
| Knowledge graph | Entity and structural relationships when the graph has been built |
| Structural view | Reconstructs heading and parent-child context at query time |
| Code graph | When available, indexes repositories, files, and symbol definitions for code-oriented exploration |
Knowledge-space configuration offers three index-method options: VectorStore, FullText, and KnowledgeGraph. Git repositories, when used as a knowledge source, support full and incremental synchronization. Once a code graph has been built for a Git repository or code files, code can be retrieved structurally by repository, file, class, and function.
Large tool results no longer have to be discarded when they exceed the inline context budget. They can be persisted and read back through a bounded file-reading tool. Citations are also carried separately from the final answer as structured data, so the frontend can render traceable source excerpts without mixing reference payloads into the answer text.
Complex tasks often contain independent branches: profile several datasets, compare multiple candidate approaches, or investigate unrelated causes before producing one conclusion. V0.8.2 lets the lead Agent delegate such branches to sub-agents and run them concurrently.
The lead Agent first records a task plan, then calls dispatch_parallel_tasks with independent work items. Each sub-agent runs with its own context, memory, conversation, and working directory. Database, knowledge-base, and read-only tool access can be inherited from the lead task, while recursive delegation is disabled.
| Capability | Description |
|---|---|
| Bounded concurrency | Runs up to three sub-agents per dispatch by default; the limit is configurable |
| Live progress | Streams running, completed, failed, and timed-out states to the frontend |
| Inspectable work | Shows each sub-agent's goal, verified steps, outputs, and artifacts |
| Final synthesis | Returns structured results to the lead Agent for one consolidated answer |
| Execution constraints | Keeps dependent work serial and prevents sub-agents from recursively delegating more work |
The per-dispatch limit can be set with service.web.agent_context.max_parallel_subagents or DBGPT_MAX_PARALLEL_SUBAGENTS. Increasing it also increases concurrent model calls and token consumption.
Parallel delegation reduces unnecessary serial waiting when work items are genuinely independent. It does not change the ordering requirements of steps that depend on one another.
Some tasks cannot be completed responsibly without a user choice: which metric definition to use, which date range applies, whether an ambiguous field should be included, or which output format is preferred. V0.8.2 adds a standard interactive question flow for these cases.
The Agent can pause execution, present one or more structured questions, and continue in the same run after the user replies. The frontend supports single choice, multiple choice, custom input, confirmation, and cancellation. Waiting is bounded so an abandoned question does not leave an execution open indefinitely.
V0.8.2 also tightens several boundaries used by Agentic workflows:
user_id and verify that resolved paths remain inside the managed upload directory.say as an argument rather than through a shell command.SKILL.md are loaded as Markdown Skills.LONGTEXT variant for large Agent messages and action reports at the ORM layer.KeyError when a chart or SQL run does not include db_name (#3199)update_flow PUT path (Fixes #3193) (#3196)LONGTEXT for large Agent messages and action reports at the ORM layer (#3189)EXAMPLE_1 database creation to the end of the schema file (#3183)user_id and constrain resolved paths for Python file uploads (Fixes #3104) (#3184)SKILL.md as Markdown Skills (#3175)inner_copy_and_install from reporting a failed build as successful (#3141)create_datasource and create_flow calls (#3138)SEEK_END position in StreamedBytesIO (#3136)VariablesProvider._convert_to_value_type (#3135)ExcelKnowledge._load (#3137)TeiRerankEmbeddings._parse_results (#3133)handleChat temporal dead zone in the frontend (#3132)This guide applies to upgrades from v0.8.1 to v0.8.2.
The V0.8.2 incremental metadata script adds one table for session- and task-scoped file persistence, three code-graph tables, the knowledge-space index-method column, and a column-width fix for Agent messages. Upgrade scripts are available under assets/schema/upgrade/v0_8_2/:
upgrade_to_v0.8.2.sql: incremental script to run on top of a v0.8.1 database.v0.8.2.sql: full V0.8.2 schema for fresh installations.As in previous releases, the incremental script targets MySQL. SQLite users should back up the metadata database before upgrading; ORM-managed tables are created when the service starts.
:::warning
To avoid data loss, back up the metadata database before upgrading. Choose the method that matches your database type, such as mysqldump for MySQL or copying the database file for SQLite.
:::
The V0.8.2 incremental upgrade applies the following metadata changes:
| Change | Description |
|---|---|
dbgpt_session_file | Stores owner-bound session and scheduled-task file metadata, stable public file IDs, managed storage URIs, inspection status, and task-file lineage. |
code_graph_vertex, code_graph_edge, code_graph_meta | Persist code-graph indexes (AST nodes, structural relationships, and per-space build metadata) for code-oriented knowledge retrieval. |
knowledge_space.index_methods | New nullable column storing the JSON list of selected index methods (for example ["VectorStore", "FullText", "KnowledgeGraph"]). |
gpts_messages.content | Widened to LONGTEXT so large Agent messages and action reports no longer fail to write. |
Apply the incremental script to your MySQL metadata database:
mysql -u <user> -p dbgpt < assets/schema/upgrade/v0_8_2/upgrade_to_v0.8.2.sql
Install or update dependencies according to your deployment method. For a source installation with the default setup:
uv sync --all-packages
Install optional extras as needed:
# Agentic Knowledge-Base Search and RAG dependencies
uv sync --all-packages --extra "rag"
# Milvus vector store
uv sync --all-packages --extra "storage_milvus"
Restart DB-GPT using your usual startup method. After startup, we recommend checking that:
Thank you to everyone who contributed to this release: @Aries-ckt, @Bartok9, @Carbene, @Dellorchid, @DreamZhongJu, @Osamaali313, @XiaoHuo888-hue, @chen-alan, @chenliang15405, @chuenchen309, @mumubuku, and @yyyCode.