Two ways to use it

Sibyl runs as an MCP server and a CLI / Python library. The key choice is who does the reasoning.

1. Retrieval provider — your model is the brain (recommended, keyless)

gather_bundle(query) searches, scrapes, dedupes, and returns structured evidence — without writing an answer. Each bundle includes versioned source/passage IDs, content hashes, retrieval timestamps, and diagnostics. gather_sources(query) runs the same retrieval and renders it as readable [Source N] blocks for conversational use. Both accept ranker="lexical" (default), ranker="flashrank" (optional local cross-encoder), or ranker="none" (preserve retrieval order). Your agent reads the evidence, cross-references it, and answers itself, citing sources and abstaining when they don't contain the answer. No API key required.

This is the highest-quality path, because a frontier model applied to real retrieved evidence beats a mid-tier model doing the synthesis and can abstain when the retrieved evidence does not contain the answer.

2. One-shot pipeline — Sibyl is the brain (needs a provider key)

research(query, depth) runs the full cycle — decompose → search → scrape → rank → synthesize → verify each finding against its source → report — using Sibyl's own configured LLM (DeepSeek by default). Use it when you want a finished report in a single call.

If no usable evidence is retrieved, or the configured LLM backend fails, this path returns an explicit failure/insufficient-evidence result instead of synthesizing from model memory.

Benchmarks

Measured on 30 questions from the official SimpleQA test set (human-verified gold, deliberately obscure long-tail facts), graded on the CORRECT / INCORRECT / NOT_ATTEMPTED rubric.

The gap isolates one variable: the model consuming the sources is the ceiling, not the sources. The same keyless retrieval that yields 17% with a mid-tier synthesizer yields ~93% when a frontier host model reasons over it. Full method, per-question results, and honest caveats: docs/EVAL_HOST_CLAUDE.md.

Reproduce the keyless path: python scripts/gather.py "<query>" and reason over the output yourself. Reproduce the one-shot path: python scripts/eval.py --depth 2 --dataset evals/gold/simpleqa_real_30.jsonl.

Quick start

MCP server (Claude Code, Cursor, …)

pip install sibyl-research

# Keyless — retrieval-provider mode, your host model reasons:
claude mcp add sibyl -- sibyl-mcp

# Or with a key, to also enable the one-shot research() tool:
claude mcp add sibyl -e DEEPSEEK_API_KEY=sk-... -- sibyl-mcp

Then, in your agent:

"Research the Serbian quarterfinalist at the 2018 Madrid Open" — uses gather_bundle or gather_sources; you synthesize

"Compare NVIDIA vs AMD vs Intel for AI workloads" — one-shot research() + compare()

CLI

pip install sibyl-research
export DEEPSEEK_API_KEY=sk-...   # or OPENAI_API_KEY, GEMINI_API_KEY, ANTHROPIC_API_KEY, …

sibyl "Canadian housing market outlook 2026"                       # standard
sibyl "Will NVIDIA keep AI-chip dominance?" -d 3 --symbols NVDA,AMD --pdf   # deep + charts
sibyl "加拿大移民政策变化" -l zh --pdf -o reports/                    # Chinese output

Tools (13 MCP tools)

gather_bundle currently returns SourceBundle schema 1.6. Its bundle_id is derived from the trimmed query, bundle status, selected URLs, and evidence hashes. Each source contains up to three passages with source-text offsets and bundle-scoped citation_id values such as sb_…/S1/P1; the combined passage text stays within chars_per_source. content_hash values are SHA-256. content_origin distinguishes direct_fetch, jina_reader, wikipedia_api, and search_snippet evidence so consumers can treat fallback snippets conservatively. When explicit publication metadata is available, published_at contains its normalized ISO 8601 value and published_at_method identifies the extraction path; this remains publisher-supplied metadata, not an independently verified date. Content-derived content_cluster_id values group exact and near-duplicate text so cross-domain syndication does not masquerade as independent evidence. relevance_score and passage score are 0–1 retrieval scores from the actual ranking backend, not probabilities or correctness judgments; they are null when ranking is disabled. Diagnostics distinguish requested_ranking_method from the actual ranking_method and expose ranking_warning when FlashRank falls back to lexical_v1.

Schema 1.6 also reports substantive_sources, independent_content_clusters, evidence_chars, evidence_sufficiency, and machine-readable sufficiency_reasons. The deterministic sufficiency check marks evidence as insufficient when there is no substantive full text, less than 200 selected evidence characters, or under 25% lexical query-term coverage. Search snippets are not substantive sources. Evidence with fewer than two substantive sources, fewer than two independent content clusters, fewer than two independent domains, or no usable lexical query terms is marked limited; limited evidence still returns bundle status ok, while insufficient evidence returns insufficient_evidence even when lead sources are included. These are retrieval-recall and independence signals, not proof that the evidence is true. quality_score remains null until a separate source-quality evaluator computes it. Check status before synthesis.

Consumers can use the SourceBundle 1.6 contract guide and checked-in JSON fixture for cross-project compatibility tests.

Within one MCP server process, matching gather_bundle and gather_sources calls share in-flight work and reuse successful evidence for 30 seconds. Failed retrievals are never cached; cached bundles retain their original retrieved_at provenance timestamps.

Offline retrieval regressions

Run the fixed, network-free ranker checks before changing retrieval scoring:

python scripts/eval_retrieval.py --ranker lexical
python scripts/eval_retrieval_pipeline.py --ranker lexical
python scripts/eval_source_quality.py
pip install 'sibyl-research[rerank]'
python scripts/eval_retrieval.py --ranker flashrank
python scripts/eval_retrieval_pipeline.py --ranker flashrank

The ranker command reports per-case first-relevant rank plus aggregate Hit@1 and MRR. The pipeline command sends the same fixed cases through search/scrape fixtures, deduplication, source and passage ranking, SourceBundle construction, and evidence-sufficiency classification; it verifies top-source accuracy, usable status, hashes, citation IDs, and source-text offsets. Both exit non-zero below their checked-in regression floors. Network I/O is replaced by deterministic fixtures, so this is a stable pipeline regression guard, not a claim about live search accuracy.

The source-quality command evaluates a deliberately limited source-type prior against contextual preference labels. It reports coverage separately from accuracy, treats tied top scores as abstentions, and includes cases where community evidence is preferable and where broad source types cannot distinguish primary from secondary reporting. This baseline is an evaluation control, not a production credibility model, so it does not populate quality_score.

How the one-shot pipeline works

You ask a question
  ├─ 1. Decompose into 3–5 focused sub-questions
  ├─ 2. Generate diverse, perspective-guided search queries
  ├─ 3. Search 4 keyless engines (DuckDuckGo, Google News, Reddit, Wikipedia; Mojeek fails over)
  ├─ 4. Scrape sources (browser headers, retry, JS-render fallback for thin pages)
  ├─ 5. Dedupe + rank by relevance
  ├─ 6. Analyze each sub-question; identify knowledge gaps → search again
  ├─ 7. Cross-reference (sentiment, consensus, disagreements)
  ├─ 8. Section-by-section synthesis (Summary, Findings, Analysis, Predictions)
  ├─ 9. Verify every finding against its cited source — flag the unsupported
  └─ Output: PDF / Markdown report with ToC, citations, charts

Depth controls cost: 1 (quick) ~20–30s · 2 (standard) ~60–90s · 3 (deep) adds gap-filling + bull/bear/base predictions.

Source reranking defaults to the dependency-free local lexical backend, so ranking does not consume an extra LLM call. Install the optional cross-encoder with pip install 'sibyl-research[rerank]', then pass ranker="flashrank" to gather_bundle / gather_sources or set reranker: flashrank for the one-shot pipeline. The model is loaded lazily and cached in-process. If FlashRank is unavailable or fails, Sibyl falls back to lexical ranking; SourceBundle diagnostics disclose that fallback. Use ranker="none" or reranker: none to preserve retrieval order. The one-shot pipeline also supports reranker: llm explicitly.

Multi-provider

Sibyl auto-detects a provider from the environment; gather_bundle and gather_sources need none.

Configure several providers with per-role routing in sibyl.yaml — e.g. cheap model for search/ranking, a stronger one for synthesis and verify:

providers:
  - model: deepseek/deepseek-v4-flash
    api_key: sk-xxx
    role: search
  - model: anthropic/claude-sonnet-4-20250514
    api_key: sk-ant-xxx
    role: synthesis

Requirements

• Python 3.10+
• gather_bundle, gather_sources, and all web search are keyless — no API keys to search the web
• URL fetching is restricted to public HTTP(S) destinations on ports 80/443; every destination and redirect is resolved, validated, and pinned to its checked public IP before TCP connection, while decompressed response bodies are capped at 2 MiB
• One LLM key only for the one-shot research() / CLI paths

License

MIT