sinapsis-bank-reconciliation 0.1.0

Sinapsis templates for bank/ERP reconciliation and LLM adjudication of fuzzy reference matches.

Sinapsis Bank Reconciliation

Sinapsis templates for bank/ERP reconciliation and LLM adjudication of fuzzy reference matches

🐍 Installation • 🚀 Templates • 📚 Usage • 🔍 License

This package provides the use-case-specific Sinapsis templates for reconciling a bank statement against an ERP/accounting export. It consumes the reference-match output produced upstream by a matching pipeline (e.g. sinapsis-polyfuzz's ExactMatch and pair matchers) and turns it into a settled, flagged reconciliation — escalating only the genuinely ambiguous fuzzy cases to an LLM adjudicator.

🐍 Installation

Install using your package manager of choice. We encourage the use of uv

Example with uv:

  uv pip install sinapsis-bank-reconciliation --extra-index-url https://pypi.sinapsis.tech

or with raw pip:

  pip install sinapsis-bank-reconciliation --extra-index-url https://pypi.sinapsis.tech

🚀 Templates

Template	Purpose
BankReconciliation	Groups rows by matched reference, compares amounts, and deterministically settles the clear-cut fuzzy cases via classify_identity, emitting reconciliation flags.
AdjudicationCaseBuilder	Turns the residual needs_human fuzzy cases into prompts for an LLM judge that decides document identity.
AdjudicationResolver	Resolves the final flags from the judge's verdicts (mirroring bank-side flags onto matched ERP rows).
ReconciliationSummaryTable	Emits a condensed, human-readable summary table of the reconciled result.

📚 Usage

These templates are referenced by class_name in a Sinapsis agent config. See the bank-reconciliation demo in sinapsis-demos for end-to-end light/heavy pipelines that wire matching → reconciliation → adjudication.

Below is a minimal example wiring the four templates this package provides. The upstream CSV reading and reference matching come from other Sinapsis packages (sinapsis-data-readers and sinapsis-polyfuzz); the values are placeholders — adapt the sources, column names and the LLM template to your data.

agent:
  name: bank_reconciliation
  description: Reconcile a bank statement against an ERP export, escalating ambiguous fuzzy matches to an LLM judge.

templates:
  - template_name: InputTemplate
    class_name: InputTemplate
    attributes: {}

  # --- Upstream: read both sources, then produce exact + fuzzy reference matches ---
  - template_name: BankCSVReader
    class_name: CSVDatasetReader
    template_input: InputTemplate
    attributes:
      root_dir: "artifacts"
      path_to_csv: "sample_bank.csv"

  - template_name: ERPCSVReader
    class_name: CSVDatasetReader
    template_input: BankCSVReader
    attributes:
      root_dir: "artifacts"
      path_to_csv: "sample_erp.csv"

  - template_name: ExactMatcher
    class_name: ExactMatch
    template_input: ERPCSVReader
    attributes:
      from_source: "BankCSVReader_x_dataset"
      from_column: "Reference1"
      to_source: "ERPCSVReader_x_dataset"
      to_column: "Reference1"
      exact_matches_source: "exact_matches"
      no_exact_matches_source: "no_exact_matches"

  - template_name: FuzzyMatcher
    class_name: RapidFuzzPairWrapper
    template_input: ExactMatcher
    attributes:
      from_source: "no_exact_matches"
      from_column: "Reference1"
      to_source: "ERPCSVReader_x_dataset"
      to_column: "Reference1"
      matches_source: "fuzzy_matches"
      min_similarity: 0.8

  - template_name: Reconcile
    class_name: BankReconciliation
    template_input: FuzzyMatcher
    attributes:
      bank_source: "BankCSVReader_x_dataset"
      erp_source: "ERPCSVReader_x_dataset"
      bank_ref_col: "Reference1"
      bank_amount_col: "Amount"
      erp_ref_col: "Reference1"
      erp_debit_col: "Debits"
      erp_credit_col: "Credits"
      match_min_similarity: 1.0
      resolve_identity: true
      overwrite: true

  - template_name: CaseBuilder
    class_name: AdjudicationCaseBuilder
    template_input: Reconcile
    attributes:
      bank_source: "BankCSVReader_x_dataset"
      bank_ref_col: "Reference1"
      prompt_path: "artifacts/adjudication_prompt.txt"

  - template_name: Judge
    class_name: LLaMACPPTextCompletion
    template_input: CaseBuilder
    attributes:
      init_args:
        llm_model_name: bartowski/Qwen2.5-3B-Instruct-GGUF
        llm_model_file: Qwen2.5-3B-Instruct-Q4_K_M.gguf
      completion_args:
        temperature: 0.1
        response_format:
          type: json_object

  - template_name: Resolver
    class_name: AdjudicationResolver
    template_input: Judge
    attributes:
      bank_source: "BankCSVReader_x_dataset"
      erp_source: "ERPCSVReader_x_dataset"
      erp_ref_col: "Reference1"

  - template_name: SummaryTable
    class_name: ReconciliationSummaryTable
    template_input: Resolver
    attributes:
      bank_source: "BankCSVReader_x_dataset"
      bank_ref_col: "Reference1"
      summary_source: "reconciliation_summary"

🔍 License

This project is licensed under the AGPLv3 license, which encourages open collaboration and sharing. For more details, please refer to the LICENSE file.

For commercial use, please refer to our official Sinapsis website for information on obtaining a commercial license.