Coverage for src/local_deep_research/advanced_search_system/tools/fetch/prompts.py: 100%

1"""Prompts used by summary-mode fetch tools. 

2 

3Kept separate from builder code so the wording can be tuned without 

4touching the wiring. Both prompts share the same role-framing block — 

5small local models (qwen3, gpt-oss) calibrate measurably better when 

6told what's downstream of their output. 

7""" 

8 

9# Role-framing block included by every summary prompt. 

10_ROLE = ( 

11 "You are a content-extraction step inside a multi-step research agent.\n\n" 

12 "Your role in the pipeline:\n" 

13 "- The agent has already run web searches and decided this specific " 

14 "page is worth reading more carefully than its snippet allowed.\n" 

15 "- Your output is returned to the agent as a tool result. The agent — " 

16 "not you — will combine your output with other sources and write the " 

17 "final cited answer.\n" 

18 "- Therefore you are an extractor, NOT an answerer. Do not interpret, " 

19 "conclude, or compose. Just pull the relevant raw text out of the page." 

20) 

21 

22# Output rules, identical for both variants. 

23_RULES = ( 

24 "Output rules:\n" 

25 "- Output ONLY verbatim quotes from the page. Copy numbers, names, " 

26 "dates, and proper nouns exactly as written — never paraphrase facts.\n" 

27 "- One quote per line. No bullets, no numbering, no section headers.\n" 

28 "- Omit navigation, ads, cookie/subscription banners, related-article " 

29 "lists, author bios, comments, and anything off-topic.\n" 

30 "- If nothing on the page helps, reply with exactly: NOT RELEVANT\n" 

31 "- Do NOT include introductions ('Here is the relevant information:'), " 

32 "conclusions ('In summary...'), explanations of what you kept or " 

33 "skipped, or commentary on the source's quality, bias, or relevance.\n" 

34 "- Maximum 1500 characters. Quality over quantity — fewer precise " 

35 "quotes beat many borderline ones." 

36) 

37 

38 

39# Focus-only variant: the agent declares what it's looking for on this page. 

40SUMMARY_FOCUS_PROMPT = ( 

41 f"{_ROLE}\n\n" 

42 "Why this page was fetched: {focus}\n\n" 

43 "Page title: {title}\n" 

44 "Page URL: {url}\n" 

45 "Page content:\n{content}\n\n" 

46 f"{_RULES}" 

47) 

48 

49 

50# Focus + overall-query variant: agent's focus PLUS the original research 

51# question, so the extractor can disambiguate vague focus phrasings 

52# ("publication year" vs "publication year of Liepmann's Prandtl-Ring award"). 

53SUMMARY_FOCUS_QUERY_PROMPT = ( 

54 f"{_ROLE}\n\n" 

55 "Overall research question: {overall_query}\n" 

56 "Why this page was fetched: {focus}\n\n" 

57 "Page title: {title}\n" 

58 "Page URL: {url}\n" 

59 "Page content:\n{content}\n\n" 

60 f"{_RULES}" 

61)