Gym/resources_servers/finance_sec_search/configs/finance_sec_search.yaml at main · NousResearch/Gym · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
# Finance SEC search: SEC filing search with optional web search (via tavily_api_key in env.yaml)
finance_sec_search_resources_server:
  resources_servers:
    finance_sec_search:
      entrypoint: app.py
      domain: agent
      verified: false
      description: SEC EDGAR filing search for financial analysis questions
      value: Enable LLMs to search and analyze SEC filings
      cache_dir: null  # defaults to ~/.cache/nemo_gym/finance_sec_search/ if not set
      use_cache: false  # keep false for eval to fetch fresh filings, set to true for training to reuse cached filings
      # Wrapped in oc.select (Gym convention, cf. labbench2_vlm) so the env resolves
      # standalone (`gym env start` / `+dry_run`); override via env / NVFlow overlay.
      # null default => web search disabled unless a key is supplied.
      tavily_api_key: ${oc.select:tavily_api_key,null}
      retrieval_model_server:
        type: responses_api_models
        name: policy_model
      judge_model_server:
        type: responses_api_models
        name: search_judge_model
      judge_responses_create_params:
        input: []
        max_output_tokens: 8192
        reasoning:
          effort: high
      # Judge prompt: loaded from this file at startup. To override inline (e.g. from
      # NVFlow), set judge_prompt_template directly — it takes priority over the file.
      judge_prompt_template_fpath: prompt_templates/finance_sec_search_judge.yaml
      retrieval_responses_create_params:
        input: []
      judge_call_timeout: 60.0  # seconds; prevents stale connections from blocking rollout
      # Retrieval prompt: loaded from this file at startup. To override inline,
      # set retrieval_system_prompt directly — it takes priority over the file.
      retrieval_system_prompt_fpath: prompt_templates/finance_sec_search_retrieval.yaml
      # Retrieval model parameters (adjust when swapping between thinking/instruct models)
      retrieval_max_output_tokens: 8192
      retrieval_model_context_length: 131072

# Judge model server (OpenAI or local — configured via env.yaml)
search_judge_model:
  responses_api_models:
    openai_model:
      entrypoint: app.py
      # oc.select w/ defaults => env resolves standalone (dry_run/bake); overlays
      # (or env.yaml) still override by setting the referenced keys.
      openai_base_url: ${oc.select:search_judge_model_base_url,https://api.openai.com/v1}
      openai_api_key: ${oc.select:search_judge_model_api_key,unset}
      openai_model: ${oc.select:search_judge_model_name,gpt-5-mini}

finance_agent:
  responses_api_agents:
    finance_agent:
      entrypoint: app.py
      max_steps: 50
      truncate_on_overflow: false
      tool_call_timeout: 300
      resources_server:
        type: resources_servers
        name: finance_sec_search_resources_server
      model_server:
        type: responses_api_models
        name: policy_model
      datasets:
      - name: example
        type: example
        jsonl_fpath: resources_servers/finance_sec_search/data/example.jsonl