agent/tasks.py

from crewai import Task
from crewai import Agent


def build_tasks(agents: dict[str, Agent], query: str, depth: str, session_id: str) -> list[Task]:
    max_sources = {"quick": 5, "standard": 10, "deep": 15}.get(depth, 10)
    is_fast_moving = depth != "deep"

    search_task = Task(
        description=f"""
Search for high-quality sources on the following research topic:

RESEARCH TOPIC: {query}
SEARCH DEPTH: {depth} (quick=5 sources, standard=10, deep=15)

Instructions:
1. Use Tavily Web Search for general/news sources — request raw_content=True
2. Use arXiv Academic Search if the topic is scientific or technical
3. Use Semantic Scholar Search for academic papers with citation counts
4. Search with 2-3 different query variations to cover multiple angles
5. Collect up to {max_sources * 2} raw results (they will be filtered next)
6. Return ALL results as a single JSON list under the key "results"

Each result must have: url, title, content, published_date, source
        """,
        expected_output=(
            "A JSON object with key 'results' containing a list of source dicts. "
            "Each dict has: url (str), title (str), content (str, full text), "
            "published_date (str or null), source (str). Minimum 8 results."
        ),
        agent=agents["researcher"],
    )

    credibility_task = Task(
        description=f"""
Score and filter the sources collected by the Research Specialist.

Instructions:
1. Take the full results list from the previous task output
2. Call the Source Credibility Scorer tool with:
   - sources_json: the JSON string of all results
   - fast_moving_topic: {str(is_fast_moving).lower()}
3. The tool will return accepted sources (score >= 0.35) and a dropped count
4. Return the accepted sources list as-is — do not modify scores

Topic context: {query}
        """,
        expected_output=(
            "A JSON object with key 'accepted' containing scored source dicts. "
            "Each dict has: url, title, content, credibility_score (float 0-1), "
            "confidence ('high' or 'low'), published_date. Sorted by credibility_score desc."
        ),
        agent=agents["credibility_analyst"],
        context=[search_task],
    )

    rag_task = Task(
        description=f"""
Embed the accepted sources and retrieve the most relevant chunks.

SESSION_ID: {session_id}
RESEARCH QUESTION: {query}

Instructions:
1. Take the accepted sources list from the previous task
2. Call Embed Sources into Vector Store with:
   - sources_json: JSON string of the accepted sources list
   - session_id: "{session_id}"
3. After embedding completes, call Retrieve Relevant Chunks with:
   - query: "{query}"
   - session_id: "{session_id}"
   - n_results: 12
4. Return the retrieved chunks with their metadata
        """,
        expected_output=(
            "A JSON object with key 'chunks'. Each chunk has: text (str), "
            "url (str), title (str), credibility_score (float), confidence (str). "
            "Also include 'accepted_sources' list for the reference section."
        ),
        agent=agents["rag_engineer"],
        context=[credibility_task],
    )

    synthesis_task = Task(
        description=f"""
Write a comprehensive, fully cited research report on:

RESEARCH QUESTION: {query}

Use ONLY the retrieved chunks provided by the RAG Pipeline Engineer.

STRICT RULES:
1. Every factual claim MUST be followed by an inline citation [N] matching a source
2. ONLY cite sources whose chunks appear in the context — never invent a citation
3. If you cannot find evidence for a claim in the chunks, write "INSUFFICIENT EVIDENCE" instead
4. Do not use any prior knowledge — only what appears in the retrieved chunks

OUTPUT FORMAT (use exactly these headers):

## Research Question
{query}

## Key Findings
[Bullet points with inline citations [N]]

## Detailed Analysis
[Multiple paragraphs with inline citations [N]]

## Points of Uncertainty or Disagreement
[What the sources disagree on, or gaps in evidence]

## Synthesis
[Your integrated conclusion from the evidence]

## References
[Numbered list: [1] Title — URL (credibility: X.XX)]

## Confidence Assessment
[HIGH / MEDIUM / LOW / INSUFFICIENT — one paragraph explaining why]

Criteria:
- HIGH: 5+ high-credibility sources (score > 0.75), consistent findings
- MEDIUM: 3+ sources with score > 0.55, or minor source disagreement
- LOW: Fewer than 3 reliable sources, or significant disagreement
- INSUFFICIENT: Could not find adequate evidence to answer reliably
        """,
        expected_output=(
            "A complete Markdown research report with inline [N] citations, "
            "a numbered reference list, and a confidence assessment label. "
            "No fabricated facts. Every claim is backed by a chunk from the retrieved context."
        ),
        agent=agents["synthesis_analyst"],
        context=[rag_task],
    )

    return [search_task, credibility_task, rag_task, synthesis_task]