Common Diagram¶

Ниже показан реальный путь вызовов для запуска:

python main.py "RAG best practices" --provider gatellm

Предположение для этого сценария: --model не передан, поэтому модель берётся из .env через settings.DEFAULT_MODEL. В вашем логе это qwen/qwen-2.5-72b-instruct.

1. Большая карта модулей и вызовов¶

flowchart TD CMD["CLI command python main.py "RAG best practices" --provider gatellm"] subgraph ENTRY["CLI / startup"] MAIN["main.py main() -> _run(args)"] SETTINGS["config/settings.py settings singleton .env + CLI overrides"] DISPLAY["ui/display.py print_banner() print_report() print_summary()"] end subgraph AGENT["agent/"] FACTORY["llm_client.py create_llm_client()"] OAI["OpenAICompatibleClient provider='gatellm'"] TRIM["_trim_history(messages)"] REQMAP["_anthropic_messages_to_openai() _anthropic_tools_to_openai()"] RESPMAP["_openai_response_to_anthropic()"] ORCH["orchestrator.py Orchestrator.run(query)"] STATE["state.py AgentState messages / sources / step / report"] end subgraph REGISTRY["tools/registry.py"] REG["ToolRegistry"] SCHEMA["get_schemas() search_web / fetch_pages / summarize_page / write_report"] NORMALIZE["_normalize_arg_names() _coerce_args()"] end subgraph TOOLS["tools/"] SEARCH["search.py search_web()"] FETCH["fetch.py fetch_pages()"] SUMM["summarize.py summarize_page()"] REPORT["report.py write_report()"] RESULT["ReportResult title / content / sources / word_count"] end subgraph EXTERNAL["external calls"] GATE["GateLLM API POST https://gatellm.ru/v1/chat/completions"] DDG["ddgs / DDGS.text() run via asyncio.to_thread()"] WEBSEARCH["search endpoints behind ddgs Wikipedia / Yahoo / Mojeek / ..."] HTTPX["httpx.AsyncClient asyncio.gather()"] URLS["Fetched URLs ACL / arXiv / Chitika / ..."] BS4["BeautifulSoup clean text + truncate to 3000 chars"] ANTH["anthropic.AsyncAnthropic optional summarize path"] end CMD --> MAIN MAIN --> SETTINGS SETTINGS -.->|LLM_PROVIDER=gatellm DEFAULT_MODEL from .env MAX_STEPS / REQUEST_TIMEOUT| MAIN MAIN --> DISPLAY MAIN --> FACTORY FACTORY --> OAI MAIN --> ORCH ORCH <--> STATE ORCH --> REG REG --> SCHEMA ORCH -->|complete with messages tools and system prompt| OAI OAI --> TRIM OAI --> REQMAP REQMAP --> GATE GATE --> RESPMAP RESPMAP --> ORCH REG -->|dispatch tool call| NORMALIZE NORMALIZE --> SEARCH SEARCH --> DDG DDG --> WEBSEARCH SEARCH -->|search result list| REG REG -->|search results| ORCH ORCH -->|add sources from search| STATE NORMALIZE --> FETCH FETCH --> HTTPX HTTPX --> URLS URLS --> BS4 BS4 --> FETCH FETCH -->|fetched page list| REG REG -->|fetched page content| ORCH NORMALIZE -. optional .-> SUMM SUMM -. direct provider bypass .-> ANTH SUMM -->|summary text| REG REG -->|summary| ORCH NORMALIZE --> REPORT REPORT --> RESULT RESULT --> REG REG -->|ReportResult| ORCH ORCH -->|store final report in state| STATE ORCH -->|return AgentState| MAIN MAIN -->|render final output| DISPLAY classDef entry fill:#e3f2fd,stroke:#1565c0,stroke-width:1px,color:#000; classDef core fill:#e8f5e9,stroke:#2e7d32,stroke-width:1px,color:#000; classDef tool fill:#fff3e0,stroke:#ef6c00,stroke-width:1px,color:#000; classDef ext fill:#fce4ec,stroke:#ad1457,stroke-width:1px,color:#000; class MAIN,SETTINGS,DISPLAY entry; class FACTORY,OAI,TRIM,REQMAP,RESPMAP,ORCH,STATE,REG,SCHEMA,NORMALIZE core; class SEARCH,FETCH,SUMM,REPORT,RESULT tool; class GATE,DDG,WEBSEARCH,HTTPX,URLS,BS4,ANTH ext;

2. Точный сценарий для вашего лога¶

sequenceDiagram autonumber actor U as User participant CLI as main.py participant CFG as settings participant FAC as create_llm_client() participant LLM as OpenAICompatibleClient participant API as gatellm /chat/completions participant ORCH as Orchestrator participant STATE as AgentState participant REG as ToolRegistry participant SRCH as tools.search participant DDG as DDGS participant FETCH as tools.fetch participant WEB as source pages participant REP as tools.report participant UI as ui.display U->>CLI: python main.py "RAG best practices" --provider gatellm CLI->>CFG: load .env into settings CLI->>CFG: override settings.LLM_PROVIDER = "gatellm" Note over CLI,CFG: settings.DEFAULT_MODEL stays from .env qwen/qwen-2.5-72b-instruct CLI->>UI: print_banner(provider, model) CLI->>FAC: create_llm_client("gatellm", model) FAC-->>CLI: OpenAICompatibleClient(provider="gatellm") CLI->>ORCH: Orchestrator(llm).run(query) ORCH->>STATE: create AgentState(query) ORCH->>STATE: append user message "Research topic: RAG best practices" ORCH->>REG: get_schemas() Note over ORCH,API: Step 1 log: react_step step=1 ORCH->>STATE: increment_step() ORCH->>LLM: complete(messages, tools, SYSTEM_PROMPT) LLM->>LLM: _trim_history() LLM->>LLM: convert Anthropic-style messages/tools to OpenAI format LLM->>API: POST /v1/chat/completions API-->>LLM: tool_calls(search_web) LLM->>LLM: normalize response back to Anthropic-style LLM-->>ORCH: stop_reason = tool_use ORCH->>STATE: append assistant tool_use ORCH->>REG: dispatch("search_web", query, max_results=5) REG->>REG: normalize aliases and coerce args REG->>SRCH: search_web(...) SRCH->>DDG: asyncio.to_thread(DDGS.text) DDG-->>SRCH: 5 results SRCH-->>REG: list[{url,title,snippet}] REG-->>ORCH: search result payload ORCH->>STATE: add_source(...) for each result ORCH->>STATE: append user tool_result Note over ORCH,API: Step 2 log: end_turn_no_content_retry ORCH->>STATE: increment_step() ORCH->>LLM: complete(updated history) LLM->>API: POST /v1/chat/completions API-->>LLM: finish_reason = stop LLM-->>ORCH: stop_reason = end_turn, no useful report ORCH->>STATE: append assistant response ORCH->>STATE: append reminder user message "You MUST call write_report now..." Note over ORCH,API: Step 3 log: tool_dispatched tool=fetch_pages ORCH->>STATE: increment_step() ORCH->>LLM: complete(updated history with reminder) LLM->>API: POST /v1/chat/completions API-->>LLM: tool_calls(fetch_pages) LLM-->>ORCH: stop_reason = tool_use ORCH->>STATE: append assistant tool_use ORCH->>REG: dispatch("fetch_pages", urls=[...]) REG->>REG: normalize aliases and coerce args REG->>FETCH: fetch_pages(urls) FETCH->>WEB: GET 5 URLs in parallel via httpx.AsyncClient WEB-->>FETCH: HTML for ACL / arXiv / Chitika pages FETCH->>FETCH: BeautifulSoup cleanup remove noisy tags truncate to MAX_CONTENT_CHARS FETCH-->>REG: list[{url,title,content}] REG-->>ORCH: fetched page payload ORCH->>STATE: append user tool_result Note over ORCH,API: Step 4 log: end_turn_no_content_retry ORCH->>STATE: increment_step() ORCH->>LLM: complete(updated history) LLM->>API: POST /v1/chat/completions API-->>LLM: finish_reason = stop LLM-->>ORCH: stop_reason = end_turn, still no report ORCH->>STATE: append assistant response ORCH->>STATE: append reminder user message "You MUST call write_report now..." Note over ORCH,API: Step 5 log: tool_dispatched tool=write_report ORCH->>STATE: increment_step() ORCH->>LLM: complete(updated history with fetched content) LLM->>API: POST /v1/chat/completions API-->>LLM: tool_calls(write_report) LLM-->>ORCH: stop_reason = tool_use ORCH->>STATE: append assistant tool_use ORCH->>REG: dispatch("write_report", title, content, sources) REG->>REG: normalize aliases and coerce args REG->>REP: write_report(...) REP-->>REG: ReportResult REG-->>ORCH: ReportResult ORCH->>STATE: state.report = result.content ORCH->>STATE: add_source(...) from result.sources ORCH->>STATE: append user tool_result ORCH-->>CLI: return final AgentState CLI->>UI: print_report(state) CLI->>UI: print_summary(state) UI-->>U: Markdown report + Sources table

3. Как читать ваш лог¶

INFO:httpx: HTTP Request: POST https://gatellm.ru/v1/chat/completions — это OpenAICompatibleClient.complete().
INFO:primp:response: ... wikipedia / yahoo / mojeek ... — это внутренние HTTP-вызовы библиотеки ddgs, которую использует tools/search.py.
INFO:httpx: HTTP Request: GET https://... — это tools/fetch.py, который параллельно грузит страницы через httpx.AsyncClient.
WARNING:agent.orchestrator: end_turn_no_content_retry — модель завершила ход без write_report, и Orchestrator добавил в историю жёсткое reminder-сообщение.
INFO:tools.report: report_written — write_report() оформил итоговый Markdown и при необходимости добавил References.
INFO:agent.orchestrator: react_done reason=write_report — оркестратор увидел терминальный инструмент и сразу завершил цикл.

4. Что важно заметить¶

В этом конкретном прогоне фактический путь был таким: search_web -> retry -> fetch_pages -> retry -> write_report.
summarize_page есть в схеме инструментов, но в вашем логе не использовался.
summarize_page сейчас вызывает Anthropic напрямую, даже если основной провайдер в сессии — gatellm.
Источники добавляются в state.sources дважды с разных сторон: после search_web и после write_report, но AgentState.add_source() дедуплицирует их по URL.