doubleo7/src/deep_research/starter.rs

use crate::deep_research::review::{self, Review};
use crate::deep_research::tools::{FetchPage, SearchWeb};
use rig::client::{AgentClientExt, Nothing};
use rig::completion::Prompt;
use rig::providers::ollama;

/// The tool-calling research loop needs to reliably decide what to search
/// for, when a page is worth fetching, and when it has enough evidence —
/// that's a reasoning-heavy job best given to the largest local Gemma
/// variant. Turning the gathered notes into prose afterwards is comparatively
/// mechanical, so the smaller/faster variant handles that pass instead.
const RESEARCHER_MODEL: &str = "gemma4:26b";
const WRITER_MODEL: &str = "gemma4-e4b:latest";
const MAX_RESEARCH_TURNS: usize = 12;
/// Research/review rounds before giving up and writing the report from
/// whatever the last pass produced, rather than looping forever on a topic
/// the reviewer can never be satisfied with.
const MAX_RESEARCH_ROUNDS: usize = 3;

const DEFAULT_TOPIC: &str =
    "What are the latest advances in running large language models locally, on consumer hardware?";

fn initialize_observability() {
    tracing_subscriber::fmt()
        .with_env_filter(
            tracing_subscriber::EnvFilter::try_from_default_env()
                .unwrap_or_else(|_| tracing_subscriber::EnvFilter::new("info")),
        )
        .with_span_events(tracing_subscriber::fmt::format::FmtSpan::CLOSE)
        .with_writer(std::io::stderr)
        .init();
}

/// Performs deep research via a two-stage agentic flow: a tool-calling agent
/// gathers and cross-checks evidence from the web, then a second agent turns
/// those raw notes into a structured report.
pub(crate) async fn start() -> anyhow::Result<()> {
    initialize_observability();

    let topic = std::env::args()
        .nth(1)
        .unwrap_or_else(|| DEFAULT_TOPIC.to_string());

    let report = research(&topic).await?;

    println!("{report}");

    Ok(())
}

/// The least-agentic shape that fits: a plain Rust loop putting *this code*,
/// not a model, in charge of when to stop — re-running research with the
/// reviewer's feedback folded in until it approves or the round budget runs
/// out, then writing the report from whatever the last pass produced.
async fn research(topic: &str) -> anyhow::Result<String> {
    let client = ollama::Client::new(Nothing)?;

    let mut findings = String::new();
    let mut feedback: Option<Review> = None;

    for round in 1..=MAX_RESEARCH_ROUNDS {
        findings = gather_findings(&client, topic, feedback.as_ref(), round).await?;

        let review = review::review_findings(&client, topic, &findings).await?;
        let approved = review.approved;

        tracing::info!(round, approved, "review verdict");

        if approved || round == MAX_RESEARCH_ROUNDS {
            break;
        }

        feedback = Some(review);
    }

    write_report(&client, topic, &findings).await
}

/// Wraps the tool-calling research loop in its own span so it's visible as a
/// single unit in traces, distinct from the writing and review phases and
/// nesting rig's own per-turn `chat`/`execute_tool` spans underneath it.
#[tracing::instrument(skip(client, feedback), fields(gen_ai.agent.name = "researcher"))]
async fn gather_findings(
    client: &ollama::Client,
    topic: &str,
    feedback: Option<&Review>,
    round: usize,
) -> anyhow::Result<String> {
    let current_date = chrono::offset::Local::now().to_string();

    let researcher = client
        .agent(RESEARCHER_MODEL)
        .name("researcher")
        .preamble(format!(
            "You are a meticulous research assistant. Use the search_web and fetch_page tools to \
             investigate the user's topic: run several searches with varied phrasing, fetch the \
             most promising pages, and cross-check claims across at least two sources before \
             trusting them. Ensure sources are up to date: the current date is {current_date}. \
             Once you are confident you have enough evidence, stop calling tools and reply with a \
             plain-text dump of every fact you gathered, using footnote-style citations: write each \
             fact followed by a bracketed number like [1], then at the end of your reply list a \
             'Sources' section mapping each number to the exact URL it came from, one per line, e.g. \
             '[1] https://example.com/page'. Reuse the same number when multiple facts come from the \
             same URL — do not give one URL two different numbers. Also call out any open questions \
             or contradictions between sources, citing the footnotes involved. This is raw research \
             material for a writer, not a final report, so favor completeness over polish.")
                      .as_str(),
        )
        .tool(SearchWeb)
        .tool(FetchPage)
        .build();

    let task = match feedback {
        None => topic.to_string(),
        Some(review) => format!(
            "Topic: {topic}\n\n\
             You already ran a research pass on this topic. A reviewer checked it against its \
             cited sources and found it insufficient. Do more research to address the reviewer's \
             feedback, then produce an updated findings dump: carry forward what's solid, and \
             add, correct, or better-source whatever the gaps call for. Gaps include out-of-date,\
             irrelevant, or clearly wrong information. The date is {current_date}.\n\n\
             Solid findings from the last pass — keep and build on these:\n{}\n\n\
             Gaps the reviewer found — conclusions not actually backed by their source, sources \
             that don't line up with the conclusion drawn from them, or parts of the topic still \
             uncovered:\n{}",
            review.solid_findings, review.gaps
        ),
    };

    let findings = researcher
        .runner(task)
        .max_turns(MAX_RESEARCH_TURNS)
        .run()
        .await?
        .output;

    tracing::info!(round, findings = %findings, "research phase complete");

    Ok(findings)
}

#[tracing::instrument(skip(client, findings), fields(gen_ai.agent.name = "writer"))]
async fn write_report(
    client: &ollama::Client,
    topic: &str,
    findings: &str,
) -> anyhow::Result<String> {
    let writer = client
        .agent(WRITER_MODEL)
        .name("writer")
        .preamble(
            "You turn raw research notes into a clear, well-organized report for the reader. The \
             notes use footnote-style citations — a bracketed number like [1] after a fact, with a \
             Sources section mapping numbers to URLs. Preserve this scheme in your report: keep the \
             same [n] markers next to the claims they support (renumbering only if you drop unused \
             sources), and end the report with a 'Sources' section listing every footnote number \
             still in use next to its exact URL. Structure the body with headings, and call out any \
             open questions or contradictions the research turned up. Do not invent facts beyond \
             what the notes provide.",
        )
        .build();

    let report = writer
        .prompt(format!("Topic: {topic}\n\nResearch notes:\n{findings}"))
        .await?;

    Ok(report)
}