Turn lead conversations into traceable follow-up decisions.
A local toolkit for lead tracking, evidence validation, buyer research and reviewable reports.
A lead, a person and a company are different records. A promising conversation is not a verified sale. A search result is not an identity match.
Lead Evidence Workbench keeps those distinctions visible. It binds observations to exact account and inquiry keys, preserves the source and timestamp of each conclusion, and generates a follow-up dashboard and workbook that a human can review.
The public repository contains source code, synthetic examples and tests. Your customer data, research observations, browser sessions and generated reports stay in your local workspace.
| Capability | Result |
|---|---|
| Collect from your own verified account | Bounded inquiries, buyer-detail enrichment, inquiry-associated visitors and advertising opportunities |
| Import saved inquiries | An idempotent SQLite lead ledger with explicit store boundaries |
| Review conversations | Product, lead quality and communication-depth assessments backed by references |
| Reuse evidence | A rebuildable history index that detects changed files, gaps and conflicting versions |
| Organize buyer context | Separate business-model and purchase-purpose assessments, with unknowns preserved |
| Measure follow-up | First non-automatic seller response, buyer silence, conversation turns and coverage gaps |
| Share an internal report | A standalone HTML dashboard and an Excel follow-up workbook |
| Run formal buyer research | A staged workflow for immutable inputs, cited research, entity matching, risk coverage and PDF review |
The toolkit validates and presents supplied evidence. Research judgments come from a human or an explicitly chosen assistant; there is no bundled model service or automatic outreach.
Requires Node.js 24+ and Python 3.12+. The offline demo needs no account, API key or browser.
git clone https://github.com/alan113523/lead-evidence-workbench.git
cd lead-evidence-workbench
python -m venv .venvActivate the environment:
# macOS / Linux
source .venv/bin/activate# Windows PowerShell
.venv\Scripts\Activate.ps1Install Python dependencies and run the demo:
python -m pip install -r requirements.txt
npm run demo -- first-demoOpen data/first-demo/dashboard.html in your browser. The same directory contains follow-up.xlsx, the synthetic source, bound messages and calculated metrics. Use a new demo name for each run: existing outputs are preserved.
The screenshot and demo records are entirely synthetic. The current dashboard and workbook labels are Chinese; the CLI and documentation are available in English and Chinese.
The Node core uses built-in modules; no npm install is required for this demo. If your Python executable has another name, set PYTHON_EXECUTABLE explicitly; see Getting started.
flowchart LR
A[Saved inquiries and messages] --> B[Exact identity and source hashes]
B --> C[Reviewable assessments]
C --> D[Follow-up metrics]
D --> E[HTML dashboard and workbook]
B --> F[Formal buyer research]
F --> G[Cited observations and risk coverage]
G --> H[Validated reports and PDF review]
Each conclusion keeps its source, observation time and uncertainty. Excluded records remain in a separate ledger. Missing information stays unknown. A payment statement never becomes a verified receipt automatically.
After running the demo, try the local SQLite importer:
node scripts/inquiries.mjs import --file data/first-demo/inquiries-demoalpha.jsonl --store demoalpha --db data/first-demo/leads.sqliteThe importer rejects rows from a different store. For your own inputs, use the single-store JSONL contract.
The repository includes collection code, not just importers. The recommended live entry point attaches to your own verified browser using an optional, pinned Playwright Core dependency. It reads inquiries and their extra fields, can enrich buyer details, and optionally collects inquiry-associated visitors and advertising opportunities. Chat-window collection and the formal package's customer-list/detail/profile collection are also included.
Try the same collection orchestrator with invented responses, without a browser or account:
npm run demo:collect -- collector-demoFor an authorized live account, first install the optional transport and create your private configuration as described in the collector guide:
npm install --prefix .deps --no-save --no-package-lock --ignore-scripts playwright-core@1.60.0
npm run collect -- --config data/collector.json --out data/new-inquiries --from 2026-01-01 --to 2026-01-07 --chains inquiriesCollection requires an existing logged-in Chromium browser with a loopback CDP endpoint, an explicit store minisite, company ID and main-account AliId. The tool checks those identities before and after reads, does not start a browser or send buyer messages, and saves outputs only under ignored data/.
Eight lower-level browser adapters are included for host integrations: inquiry lists, member-linked inquiries, buyer levels, exact CRM links, order candidates, assignment/pool records, visitors and advertising reports. They require fresh store/route proofs and host verification before accepting results; their scopes and entry points are listed in the collector inventory.
Offline tests and the synthetic demo do not certify access to a real account or complete platform coverage. Account IDs, session artifacts and live results are never bundled.
The optional formal-review/ package includes input validation, browser-collection adapters, cited research records, entity matching, nine-dimension analysis, risk checks, report generation and PDF validation.
Try an entirely offline workflow with no matched entities:
python -X utf8 formal-review/examples/synthetic_demo.py --out data/formal-demo --customers 3 --matched 0For a real task, supply an authorized JSON/XLSX pair and a new workspace, then follow the next instruction at each stage:
python -X utf8 formal-review/scripts/workflow.py init --input data/input.json --xlsx data/input.xlsx --workspace data/research
python -X utf8 formal-review/scripts/workflow.py next --run data/research/RUN_DIRECTORYReal research requires an operator and accessible sources. Browser collection requires your own logged-in, explicitly verified account. PDF rendering requires a local Chrome/Edge installation and appropriate fonts. A PDF completion receipt also requires human page inspection; setting a flag does not bypass the validators. See the formal workflow guide.
- No real customer datasets, account identifiers, browser profiles or company operating reports are distributed.
- Runtime records and reports belong under Git-ignored
data/; generated files are private by default. - Store keys are caller-supplied.
demoalpha,demobetaanddemogammaare invented examples. - The public project starts with a fresh history. Internal commits, handoffs, screenshots and frozen company runs are excluded.
- The ordinary demo performs no network calls. Optional browser and public-research workflows require explicit operator action.
Read the privacy model before importing real business data. A public source repository does not make your generated dashboard safe to publish.
npm test # Node, Python core and formal workflow tests
npm run test:node
npm run test:python
npm run test:formal
npm run check:privacy # Tracked-path and personal-machine-path checksTests use temporary synthetic fixtures. Platform-specific integrations and real browser/account flows are separate from the offline test suite. Windows runners use explicit UTF-8 Python mode.
- Getting started and input examples
- Collectors, private configuration and module inventory
- Architecture and evidence boundaries
- Privacy and publication policy
- Formal buyer-review workflow
- Release verification
- Contributing · Security policy · Changelog
This is an early public toolkit, not a hosted CRM service. It does not send buyer messages, write CRM records, confirm bank receipts or make legal/credit decisions. Internal frozen-run compatibility is deliberately excluded: public evidence runs must pin the published source version.
The public layout and onboarding were informed by Datasette, Twenty and Playwright: a clear purpose, runnable example, visible limitations and a direct path to development docs. Their branding and product code are not bundled.
MIT. Retained component copyrights and template icon notices are listed in NOTICE.md and the formal template notices. Browser binaries, fonts and external sanctions datasets are not distributed.
