Smith Research — Website & Company Deep Research

Produce an auditable research dossier on a target site + company. The design separates extraction (deterministic scripts, zero LLM) from interpretation (LLM subagents that see ONLY already-stored, cited chunks). The guarantee is traceability, not omniscience: every published sentence is citation-gated and adversarially entailment-checked, and a deterministic LLM-free gate fails the run rather than ship an uncited claim.

Arguments: $ARGUMENTS (first positional token is the target domain)

When to use

• You need a complete, source-backed picture of a company: what it is, what it offers, pricing, target audience, marketing angle, plus background — PR/news, funding, and leadership.
• You want the underlying corpus stored so you can ask follow-up questions later.

When NOT to use

• Quick one-off lookups (use a direct web search).
• Authenticated/paywalled content (out of scope).

Vault Logging

Throughout, log significant events to the vault session log. Read the path from .smith/vault/.current-session. If missing or the vault is not initialized, skip logging silently. Immediately before every Agent tool call, append a block naming the phase, subagent_type, and model (the Agent return value does not expose these to the parent, so this is the only capture point).

Prerequisites (checked in Phase 0)

• Docker running with a Qdrant container on :6333.
• Ollama running with nomic-embed-text pulled.
• The skill-owned venv + Playwright browser + Playwright WS server (Phase 0 bootstraps these).

Engine (deterministic CLI — no LLM)

All crawling/extraction/indexing/retrieval/verification bookkeeping is done by:

python3 skills/smith-research/scripts/run.py <subcommand> [flags]

Subcommands: check, discover, crawl, index, background, retrieve, evidence-check, record-claim, verify-report, status. See skills/smith-research/scripts/README.md and specs/32-smith-research/contracts/cli.md.

The LLM only ever runs in Phases 5–6 below, over retrieved+cited chunks.

---

Phase 0: Bootstrap & prerequisites

SR=skills/smith-research/scripts
# 1. venv (idempotent)
[ -d "$SR/.venv" ] || python3 -m venv "$SR/.venv"
source "$SR/.venv/bin/activate"
pip install -q -r "$SR/requirements.txt"
# 2. Playwright browser (required for screenshots + render fallback + WS server)
python -m playwright install chromium >/dev/null 2>&1
# 3. host-side Playwright WS server on :9224 (idempotent; uses venv python)
bash "$SR/start-playwright-server.sh"
# 4. verify Qdrant + Ollama + Playwright are all reachable
python "$SR/run.py" check --domain "$DOMAIN" --json

If check exits non-zero (2), STOP and surface the remediation lines it printed (start Qdrant / ollama pull nomic-embed-text / start the WS server). Do not proceed — correctness depends on all three being up.

macOS note (S5): artifacts write to ~/Documents/smith-research/, which may trigger a one-time "allow Terminal to access Documents" TCC prompt, and syncs to iCloud by default — crawled pages/SERP captures may contain PII. Warn the user once. Set SMITH_RESEARCH_OUTPUT_DIR to override the location.

Phase 1: Discover

python "$SR/run.py" discover --domain "$DOMAIN" --json

Builds the URL frontier (sitemap recursion → robots → BFS fallback). Note the run_dir it prints — every later phase resolves the latest run automatically. If discovered is 0, STOP (exit 3): the domain has no reachable pages.

Phase 2: Crawl

python "$SR/run.py" crawl --domain "$DOMAIN" ${RESUME:+--resume} --json

Fetches every URL (static → render fallback), saves markdown + raw HTML + screenshot, checkpoints, and enforces terminal statuses so the run always completes. A non-zero exit (4) means some URLs reached a terminal-unsuccessful state (blocked/failed) — that is an honest partial, NOT a hard failure; note the counts and continue.

Phase 3: Index

python "$SR/run.py" index --domain "$DOMAIN" ${RESUME:+--resume} --json

Chunks + embeds crawled pages into the per-run _site Qdrant collection (idempotent — safe to re-run).

Phase 4: Background research

python "$SR/run.py" background --domain "$DOMAIN" --depth "${DEPTH:-thorough}" --json

Runs Bing-primary SERP + Reddit across the research topics (demographic, marketing, pr, funding, leadership, competitors, sentiment), assigns a source_tier per source, deep-fetches and embeds into the _background collection. Blocked/empty engines are logged as gaps, never fabricated. Note per_tier — social_unverified sources must be hedged in synthesis.

Phase 5: Synthesize (LLM — grounded, cited)

Write report.md progressively, ONE section at a time, each by a subagent that sees ONLY retrieved chunks. Report sections (fixed order): Company · Offerings · Pricing · ICP/Audience · Marketing Angle · News/PR · Funding · Leadership.

For each section:

1. Retrieve evidence (site for 1–5, background for 4–8; both allowed):

   python "$SR/run.py" retrieve --domain "$DOMAIN" --query "<section-focused query>" --collection both --k 10 --json
   

2. Spawn ONE general-purpose subagent (log the Agent call per Vault Logging). Pass it ONLY the retrieved chunks (text + url + source_tier). Prompt contract:

   "Write the section using ONLY the provided sources. If a fact is not in the sources, do NOT state it. Cite every factual sentence inline as [<url>]. For any source marked social_unverified, attribute and hedge ('Per a Reddit thread, …') — never state it as bare fact. Return markdown."

3. Append the returned section to report.md.

Run sections in parallel where practical (they are independent). Update run.json/run_state phase to synthesize with per-section status so a died run resumes at the next unfinished section (S3).

Phase 6: Verify (LLM per-claim) + DETERMINISTIC gate

6a — Adversarial per-claim verification. For each section's draft:

1. Decompose it into ATOMIC claims (one assertion each — split compound sentences so no fact rides along uncounted).
2. For each atomic claim, gather candidate evidence:

   python "$SR/run.py" evidence-check --domain "$DOMAIN" --claim "<atomic claim>" --json
   

3. Spawn a verifier general-purpose subagent (adversarial prompt):

   "Default to UNVERIFIED. Mark supported ONLY if a provided chunk directly ENTAILS the claim. Mark contradicted if a chunk refutes it. Topical similarity is NOT support. Return the verdict + which chunk_id(s) support it."

4. Record every claim's outcome:

   python "$SR/run.py" record-claim --domain "$DOMAIN" --section "<S>" --claim "<c>" \
     --verdict supported --source-urls <urls...> --chunk-ids <ids...> \
     --chunk-texts "<supporting chunk text...>" --source-tier <tier>
   

5. In report.md, strip any claim that came back contradicted, and mark any unverified claim with a trailing [UNVERIFIED] or remove it.

6b — Deterministic closing gate (LLM-FREE — the actual enforcement):

python "$SR/run.py" verify-report --domain "$DOMAIN" --json

This parses the FINAL report.md and asserts every declarative sentence maps to a verdict=supported claim with a resolvable chunk. If it exits 4, it prints the offending sentences (uncited / cited-but-not-supported). You MUST fix each (add a real citation from evidence, or remove the sentence) and re-run until it exits 0. Do not present the report to the user until this gate passes. This is the guarantee: the shipped report contains zero uncited claims.

Phase 7: Finalize

• Confirm verify-report exited 0.
• Report to the user: the run_dir path, report.md, evidence.jsonl, the crawl/background counts, and any honest gaps (blocked engines, terminal URLs, [UNVERIFIED] items).
• Tear down the Playwright WS server if this run started it:

  kill "$(cat "${TMPDIR:-/tmp}"/smith-research-playwright-9224.pid 2>/dev/null)" 2>/dev/null || true
  

Later: ask questions of the stored corpus (US-3)

The vector collections persist. To answer follow-ups from stored evidence:

python "$SR/run.py" retrieve --domain "$DOMAIN" --query "<question>" --collection both --k 8 --json

Answer ONLY from returned chunks + their URLs; if a topic is absent, say so — never fill the gap from model memory.

Key rules

• Extraction is never done by an LLM. Phases 1–4 and all bookkeeping are the deterministic CLI.
• Synthesis sees only retrieved, cited chunks. No model memory, ever.
• The deterministic gate is the guarantee. A run that can't pass verify-report is not finished — fix or cut the offending claims.
• social_unverified is never bare fact. Hedge and attribute.
• Never fabricate around a blocked source. Log the gap; report it honestly.