Bank-to-Ledger Matching Workflow: OpenAI Embeddings, Slack, Postgres

Bank-to-Ledger Matching with OpenAI Embeddings + Slack Human Approval

Automatically match bank transactions to ledger entries using fuzzy scoring and optional OpenAI-compatible embeddings—then route ambiguous cases to Slack for HMAC-signed, expiring human approval. Results, review queues, and embedding cache are stored in PostgreSQL as the workflow runs.

What this workflow does

• Triggers: Run manually, on a 4-hour schedule, or via a POST webhook.
• State + security: Loads prior matching state and an approval secret from PostgreSQL.
• Fetch & normalize: Pulls two record sets from configured HTTP APIs, normalizes and de-duplicates them, and stops if either source fails.
• Candidate pairing: Builds candidate pairs using amount/date blocking, then computes fuzzy similarity features across amount, date, counterparty, description, references, and learned aliases.
• Optional embeddings: Loads cached vectors from PostgreSQL and calls an OpenAI-compatible embeddings endpoint only for missing texts. Semantic similarity is fused into the text score without overriding numeric mismatches.
• One-to-one matching: Automatically matches only high-confidence, non-ambiguous pairs; creates a review queue for the rest with top candidates and reasons.
• Slack approval: Saves run summaries, auto-matches, review items, and any new embedding vectors to PostgreSQL, then posts review requests to Slack with HMAC-signed, expiring approval links.
• Decision handling: Serves a GET review page and a POST decision endpoint to apply reviewer approval/rejection in PostgreSQL and update learned aliases.

Use cases

• Reduce manual reconciliation by linking bank transactions to ledger entries with a confidence-based review loop.
• Improve matching quality over time using learned aliases and cached embeddings.
• Operationalize human-in-the-loop approvals for ambiguous matches in Slack.

Technical details

• Integrations: Slack, PostgreSQL, configurable HTTP APIs, and an OpenAI-compatible embeddings endpoint.
• n8n nodes used: if, code, no op, merge, webhook, postgres.