Repository navigation
Conversation
…'s main openai was pinned to 2.44.0 to match nemo-gym 0.6.0. nemo-gym's main pins 2.45.0, so a project that installs both cannot resolve. Accept a range, >=2.44.0,<3, and lift langchain-openai's <1.3.5 cap, which only followed the pin. The lock keeps openai 2.44.0 and langchain-openai 1.3.4. Signed-off-by: Felipe Vieira Frujeri <ffrujeri@nvidia.com>
|
No actionable comments were generated in the recent review. Walkthrough
Merge Risk:
|
…t Server, on BIRD dev Auto Ontology's main answers a question with one ReAct episode (react.loop.run_episode): its model searches Auto Ontology's catalog, runs SQL and submits a query. The agent calls run_episode as Auto Ontology's own ReAct step does, with auto-ontology-eval's settings for a graded benchmark, and answers with the submitted query in a fenced sql block. - The task (question, evidence, database) comes in the seed's agent context. birdbench's prepare.py writes it with +prepare_script_args.agent_context=true, to its own file, and benchmarks/birdbench/auto_ontology.yaml pairs the agent with bird_sql through single_agent_turn_legacy. - The episode's model is Auto Ontology's reasoning client on the policy's Model Server. Its calls go through Gym's ServerClient on the activation's rollout route, with the invocation ID, and report their usage (model_calls.py). - Auto Ontology is pinned to its main (a630ebd) plus the commit of NVIDIA/auto-ontology#543 that accepts the openai version Gym pins. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Signed-off-by: Felipe Vieira Frujeri <ffrujeri@nvidia.com>
…t Server, on BIRD dev Auto Ontology's main answers a question with one ReAct episode (react.loop.run_episode): its model searches Auto Ontology's catalog, runs SQL and submits a query. The agent calls run_episode as Auto Ontology's own ReAct step does, with auto-ontology-eval's settings for a graded benchmark, and answers with the submitted query in a fenced sql block. - The task (question, evidence, database) comes in the seed's agent context. birdbench's prepare.py writes it with +prepare_script_args.agent_context=true, to its own file, and benchmarks/birdbench/auto_ontology.yaml pairs the agent with bird_sql through single_agent_turn. - The episode's model is Auto Ontology's reasoning client on the policy's Model Server. Its calls go through Gym's ServerClient on the activation's rollout route, with the invocation ID, and report their usage (model_calls.py). - Auto Ontology is pinned to its main (a630ebd) plus the commit of NVIDIA/auto-ontology#543 that accepts the openai version Gym pins. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Signed-off-by: Felipe Vieira Frujeri <ffrujeri@nvidia.com>
… once Pin Auto Ontology once, in benchmarks/birdbench/auto_ontology/pyproject.toml, and build everything Auto Ontology-based from that pin. - The catalog image (auto_ontology/Dockerfile, build_image.sh): Auto Ontology at the pinned commit from its own repository and lock file, the BIRD dev databases, and a catalog Postgres. build_catalog.py migrates the catalog with Auto Ontology's Alembic migrations, then for each database ingests the schema, writes BIRD's column descriptions from its database_description CSVs (Auto Ontology has no input for them, and a SQLite database carries none) through apply_metadata_batch, embeds the described catalog and compiles the semantic layer. The build fails when a model call failed after Auto Ontology's retries or a database has no vectors; Auto Ontology logs both and goes on. Keys reach the build as a BuildKit secret. - The catalog-tools rows now start as Auto Ontology's agent starts an episode: its system prompt with the SQLite rules, the question with the evidence apart, and its tool definitions, from the pinned package (auto_ontology/react_prompt.py). The auto_ontology project prepares them; Gym's own prompt and tool descriptions go. - prepare.py --agent-context adds the task to each row's agent context, for the Auto Ontology agent. - The catalog server's defaults follow the image; its routes pass any argument on, so Auto Ontology answers a wrong one as its loop does. The pin is Auto Ontology's main (a630ebd) plus the commit of NVIDIA/auto-ontology#543, which accepts the openai version Gym pins. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Signed-off-by: Felipe Vieira Frujeri <ffrujeri@nvidia.com>
What changed
pyproject.tomlnow accepts a range ofopenaiversions instead of one exact version:openai==2.44.0>=2.44.0,<3langchain-openai>=1.3.0,<1.3.5>=1.3.0uv.lockchanges only these two specifiers. It still installs openai 2.44.0 and langchain-openai 1.3.4, so this repository's own environment does not change.Why
The
openaipin was chosen to match the version nemo-gym 0.6.0 declares, and thelangchain-openaicap only followed that pin (releases after 1.3.4 needopenai>=2.45.0). nemo-gym's main now pinsopenai==2.45.0, so a project that installsauto-ontology-servernext to it cannot resolve.That project is an Auto Ontology Agent Server in NeMo Gym. It installs this package from
mainand calls the ReAct agent'srun_episodewith a model client that goes through Gym's Model Server, so each model call is captured per rollout for training. With a range, both nemo-gym 0.7.0 (openai==2.44.0, which auto-ontology-eval uses) and nemo-gym's main (openai==2.45.0) resolve, and the next Gym bump does not break the install again.No issue was filed for this; the change is the whole request.
How it was tested
Following CONTRIBUTING.md (Postgres with pgvector in Docker, fixtures seeded,
alembic upgrade head):uv lock --checkruff checkandruff format --checkonauto_ontology alembic dev_toolsuv run pytest auto_ontology dev_tools -qwith the locked openai 2.44.0 and langchain-openai 1.3.4uv run --with ...)cd mcp && uv run pytest -qSummary by CodeRabbit