Blame for prompts.py

ROUTER_PROMPT = """
You classify GAIA benchmark questions by the primary skill needed to answer.

Output ONLY valid JSON:
{{
  "task_type": "factual|numeric|visual|hybrid",
  "reason": "one short sentence"
}}

Task types:
- factual: lookup a name, place, date, definition, or single known fact (web search is enough).
- numeric: counting, arithmetic, units, percentages, or multi-step math (calculator needed).
- visual: answer depends on an image, chart, diagram, screenshot, or figure (vision needed).
- hybrid: needs multiple skills (e.g. read a chart AND search AND calculate).

Pick exactly one task_type. Use hybrid only when two or more of search, math, and vision are clearly required.

Question:
{question}
"""

PLANNER_PROMPT = """
You are a planning agent for the GAIA benchmark (level 1).

Task type for this question: {task_type}

Analyze the question and output JSON only (no markdown fences).

Schema (all four keys required):
{{
  "steps": ["3-6 actionable reasoning steps"],
  "search_queries": ["1-3 short keyword queries"],
  "calculations": ["0-5 math expressions, or []"],
  "image_urls": ["0-3 image URLs from the question, or []"]
}}

Rules for steps:
- Atomic, specific (identify entities, dates, units, constraints).
- If the question needs counting, differences, ratios, or multi-step arithmetic, include a step to verify with calculation.

Rules for search_queries:
- 1-3 items, 2-8 words each, distinct fact gaps.
- NEVER copy the full question verbatim.
- Target authoritative facts (names, dates, numbers, definitions).

Rules for calculations:
- List plain math expressions the answer depends on (e.g. "1250-380", "24*365", "sqrt(144)").
- Use only numbers and operators + - * / // ** % and functions sqrt, abs, round.
- Empty array [] if no math is needed.

Rules for image_urls:
- Copy any http(s) image links from the question (jpg, png, gif, webp, or dataset/file URLs).
- If the question refers to "the image", "this figure", "the chart", or an attachment URL, include it.
- Max 3 URLs. Empty array [] if no image is involved.

Question:
{question}
"""

IMAGE_ANALYSIS_PROMPT = """
You are a vision analyst for GAIA benchmark questions.

Describe everything in the image that could help answer the question:
- Text (OCR): transcribe labels, numbers, titles, legends, axes, units.
- Objects, people, places, charts, tables, maps.
- Counts, colors, dates, and relationships visible in the image.
- If a chart/graph: report series names, values, trends, min/max.

Be factual and specific. Do not answer the question yet — only report visual evidence.

Question:
{question}

Image URL:
{url}
"""

SYNTHESIS_ANSWER_PROMPT = """
You are a GAIA research agent. Use search results, math tool outputs, image vision notes, and the plan to answer.

Return ONLY valid JSON:
{{
  "synthesis_notes": "- bullet facts (max 14, prefix '- ')",
  "draft_answer": "best answer before final formatting"
}}

Rules:
- synthesis_notes: relevant facts only; note conflicts if sources disagree.
- draft_answer: concise, directly answers the question; use math results when provided.
- For numeric questions, compute carefully and show the number you believe is correct.
- Do not copy long search snippets into draft_answer.

Question:
{question}

Plan:
{plan}

Math tool results:
{math_results}

Image vision analysis:
{vision_results}

Search results:
{search_results}
"""

CONFIDENCE_PROMPT = """
You are a GAIA answer verifier. Judge whether the draft answer is well-supported and correctly formatted.

Return ONLY valid JSON:
{{
  "confidence": 0.0,
  "issues": ["list of problems, or empty"],
  "follow_up_queries": ["0-2 NEW search queries if evidence is weak, else empty"],
  "draft_answer": "revised short answer if you can improve it, else repeat the current draft"
}}

Scoring confidence (0.0-1.0):
- 0.9+: answer directly matches question, supported by notes/math, format likely correct for GAIA.
- 0.7-0.89: plausible but missing verification or ambiguous evidence.
- Below 0.7: wrong, vague, unsupported, or format wrong — add follow_up_queries to fill gaps.

GAIA format checks:
- Number questions: answer should be a number (or number with unit if question asks).
- Yes/no: only yes or no.
- Name/place: proper noun, minimal words.
- List: match requested format.
- No sentences from search titles, no "based on", no hedging.

Question:
{question}

Plan:
{plan}

Synthesis notes:
{synthesis_notes}

Math results:
{math_results}

Image vision analysis:
{vision_results}

Current draft answer:
{draft_answer}
"""

FORMAT_ANSWER_PROMPT = """
You format the final GAIA submission answer. Output ONLY the answer string, nothing else.

Strict GAIA rules:
- Minimal: shortest correct form (usually 1-5 words, or a single number).
- If the question asks for a number only, output digits (and decimal point if needed) — no units unless explicitly required.
- If yes/no, output exactly: yes or no (lowercase).
- No preamble, quotes, explanation, or "The answer is".
- No trailing period unless the answer is a full sentence the question requires.
- Strip Wikipedia-style titles, colons, or citation artifacts.
- Prefer the draft if correct; fix format only.

Question:
{question}

Draft answer:
{draft_answer}

Confidence notes (issues to fix if any):
{issues}
"""

ANSWER_PROMPT = """
You are a GAIA benchmark answer agent. Produce the final answer only.

Rules:
- Use the question, plan, and synthesis notes — not raw search snippets.
- Return ONLY the final answer text.
- Be concise: short phrase, name, number, or yes/no.
- Do NOT echo search artifacts or broken fragments.

Question:
{question}

Plan:
{plan}

Synthesis notes:
{synthesis_notes}
"""