backend/frame/prompts/udr_minimal_generating/2.routine_generating.txt

1. Send a notification of type "prompt_received" with description saying what PROMPT has been received, e.g. "Received research request: {PROMPT}"
2. Send a notification of type "prompt_analysis_started", with description indicating that we are now analyzing the research request.
3. Take the PROMPT and ask a language model to produce 3 search phrases that could help with retrieving results from search engine for the purpose of compiling a report the user asks for in the PROMPT. The search phrases should be simple and objective, e.g. "important events 1972" or "energy consumption composition in India today". Use a long prompt for the model that describes in detail what is supposed to be performed and the expected output format. Instruct the model to return the search phrases on one line each. Tell the model not to output any other text -- just the newline-separated phrases. Then, parse the output of the language model line by line and save the resulting search phrases as "phrases" for further research, skipping over empty lines.
4. Send a notification of type "prompt_analysis_completed", with a description saying as much.
4.1 Send a notification of type "task_analysis_completed", informing the user that the search plan has been completed and informing them how many search phrases will be invoked, e.g.  "Search planning completed. Will be searching through {len(topics)}+ terms."
5. For each phrase in phrases output by step 3., perform the following:
    - Send a notification of type "search_started", with the description indicating what search phrase we are using for the search, e.g. "Searching for phrase '{phrase}'"
    - Perform search with the phrase.
    - Once the search returns some results, append their contents to CONTEXT one by one, separating them by double newlines from what is already present in the CONTEXT.
    - Send a notification of type "search_result_processing_completed", indicating in its description that the search results for term {term} have been processed.
6. Send a notification to the user with type "research_completed", indicating that the "Research phase is now completed.".
7. Send a notification with type "report_building", with the description indicating that the report is being built.
8. Take CONTEXT. Call the language model, instructing it to take CONTEXT (to be appended into the LM call) and produce a deep research report on the topic requested in PROMPT. The resulting report should go into detail wherever possible, rely only on the information available in CONTEXT, address the instruction given in the PROMPT, and be formatted in Markdown. This is to be communicated in the prompt. Do not shy away from using long, detailed and descriptive prompts! Tell the model not to output any other text, just the report. The result produced by the language model is to be called REPORT.
9. Send a notification with type "report_done", indicating that the report has been completed. Add "report" as a field containing the REPORT to be an additional payload to the notification.
10. Output the REPORT.