ds-caselaw-data-enrichment-service icon indicating copy to clipboard operation
ds-caselaw-data-enrichment-service copied to clipboard

Judgment citation annotations for the National Archives Find Case Law service

The National Archives: Find Case Law

This repository is part of the Find Case Law project at The National Archives. For more information on the project, check the documentation.

Judgment Enrichment Pipeline

Tests Coverage Maintainability

Mark up judgments in Find Case Law with references to other cases and legislation.

1. Introduction

This resource documents the design and operation of the Judgment Enrichment Pipeline (JEP) built for The National Archives by MDRxTECH and vLex Justis to support the publishing process that sits behind the Find Case Law platform.

The primary purpose of the JEP is to "enrich" the judgments published on Find Case Law by marking up important pieces of legal information - such as references to earlier cases and legislation - cited in the body of the judgment. In certain scenarios described elsewhere in this documentation, the JEP will "repair" or resolve entities that are malformed whilst respecting the original text of the judgment in question.

1.1 The general anatomy of the JEP

At its core, the JEP is a series of serverless functions, which we call Annotators, that sequentially add layers of markup to judgments submitted for enrichment. Each Annotator is responsible for performing a specific type of enrichment. For example, the Case Law Annotator detects references to case law citations (such as [2021] 1 WLR 1) and the Legislation Annotator is responsible for marking up mentions of UK primary legislation. An overview of the Annotators can be found below with more detailed notes on each set out in dedicated documentation in this folder.

A comprehensive map of the JEP's architecture can be found here

The Annotators are supported by a cast of utility functions that are responsible for ETL, XML validation, rules and data management and file manipulation. The most important of these utility functions are the Replacers, which generate the enriched XML that is sent back for publication on Find Case Law.

A significant amount of core markup annotation is provided directly by the JEP, but it is also supported an integration with the vLex vCite engine. vCite extends the JEP's functionality in a range of ways, including the addition of a comprehensive suite of case law citation matchers. See here for more detail on the vCite integration and how it is controlled.

Phases of Enrichment

An example enriched snippet of LegalDocML feature case law citation markup looks like this:

<ref href="https://caselaw.nationalarchives.gov.uk/ewca/civ/2021/1308" uk:canonical="[2021] EWCA Civ 1308" uk:isneutral="true" uk:type="case" uk:year="2021" uk:origin="TNA">[2021] EWCA Civ 1308</ref>, <ref href="#" uk:canonical="[2022] 1 WLR 1585" uk:isneutral="false" uk:type="case" uk:year="2022" uk:origin="TNA">[2022] 1 WLR 1585</ref>

1.2 The Annotators

The JEP is a modular system comprising a series of AWS Lambda functions -- the Annotators -- that are each responsible for performing a discrete step in the enrichment pipeline. The five Annotator functions are:

  1. Case Law Annotator -- detects references to UK case law citations, such as [2022] 1 WLR 123
  2. Legislation Annotator -- detects references to UK primary legislation, such as Theft Act 1968
  3. Abbreviation Annotator -- detects abbreviations and resolves them to their longform. For example, the longform of HRA 1998 is Human Rights Act 1998
  4. Oblique Legislative References Annotator -- detects indirect references to primary legislaton, such as the Act or the 1998 Act and determines which cited primary enactment the indirect reference corresponds to
  5. Legislative Provision Annotator -- identifies references to legislation provisions, such as section 6, and identifies the corresponding primary enactment, for example section 6 of the Human Rights Act

1.3 Enrichment phases

There are four phases of enrichment. Each phase of enrichment generates enriched LegalDocML that is progressively enriched by each successive phase of enrichment.

First Phase Enrichment The first phase of enrichment consists of the Case Law Annotator; the Legislation Annotator; and the Abbreviations Annotator

Second Phase Enrichment The second phase of enrichment consists of the Oblique Legislative References Annotator

Third Phase Enrichment The third phase of enrichment consists of the Legislative Provision Annotator

Fourth Phase Enrichment The fourth and final phase of enrichment consists of the vCite integration

1.4 Replacers

The Replacers are responsible for registering the various entities detected by the Annotators, including their entity types and position in the judgment body. The registered replacements are then applied to the judgment body through a series of string manipulations by the make_replacements lambda.

There are two sets of replacer logic. The first set provides the logic for first phase enrichment replacements. The second set of replacer logic handles replacement in the second and third phases of enrichment.

1.5 Re-enrichment

It is possible for the same judgment to be submitted for enrichment on multiple occasions, which creates the risk that existing enrichment present in the judgment will break as additional enrichment is added to the judgment. To address this, the JEP "sanitises" the judgment body prior to making replacements. The sanitisation process is simply performed by stripping existing </ref> tags from the judgment. This logic is handled in the make_replacements lambda.

IMPORTANT: the sanitisation step does not currently distinguish between enrichment supplied by the JEP itself, by vCite or from some other source! Particular care should be taken to avoid inadvertently removing vCite enrichment by re-enriching a judgment that includes vCite enrichment when the vCite integration is switched off.

2 Adding new citation rules to the Case Law Annotator

The Case Law Annotator uses a rules-based engine, the Rules Manifest, which is built on top of the spaCy EntityRuler to detect case law citations (e.g. `[2022] 1 WLR 123). The Rules Manifest is stored as a table in Postgres where each row in the table represents a rule.

The creation of rules is currently managed by modifying and uploading a CSV version of the Rules Manifest, which is stored in production-tna-s3-tna-sg-rules-bucket with a filename conforming to the pattern yyyy_mm_dd_Citation_Manifest.csv.

See here for guidance on how to create and modify rules in the Rules Manifest.

3 Enriching judgments: How to run the pipeline

There are two ways to operate the pipeline:

  1. Triggering the pipeline via file upload to S3
  2. API integration with the MarkLogic database

3.1 Triggering the pipeline via file upload to S3

3.1.1 Upload the judgment XML to the origin bucket in S3

The JEP can be operated manually by uploading judgments directly to the JEP's trigger S3 bucket: s3://production-tna-s3-tna-sg-xml-original-bucket/. We recommend using the AWS CLI to achieve this, like so:

aws s3 cp path/to/judgment.xml s3://production-tna-s3-tna-sg-xml-original-bucket/

3.1.2 Collect the enriched XML file from the terminal bucket in S3

The enrichment process typically takes five-six minutes per judgment. Enriched judgment XML is deposited in the JEP's terminal bucket: s3://production-tna-s3-tna-sg-xml-third-phase-enriched-bucket. Again, we recommend using the AWS CLI to retrieve the enriched XML, like so:

aws s3 cp s3://production-tna-s3-tna-sg-xml-third-phase-enriched-bucket/judgment.xml path/to/local/dir

3.2 API integration with the MarkLogic database

The standard mechanism for triggering the enrichment pipeline is via the TNA editor interface.

4 Tests

There is a suite of tests that can be run locally with pytest -m "not integration" or scripts/test but you'll need to ensure you've installed src/tests/requirements.txt

You can also obtain a test coverage report with coverage run --source . -m pytest && coverage report

The tests are currently run in CI as specified in .github/workflows/ci_lint_and_test.yml

5 Architecture

Architecture The VCite integration is shown more distinctly in the diagram below: VCite-integration

6 Workflow

CI/CD works in the following way:

  • Engineer branches from main branch, commits code and raises a pull request.
    • The code within the repo is checked using the tools defined in .pre-commit-config.yaml.
    • The Terraform code is checked against a linting tool called TFLint.
    • Terraform is validated and planned against staging and production as independent checks.
  • Upon merge, non dockerised lambdas are built, terraform is planned, applied and then docker images are built and pushed to ECR. This occurs for staging, if staging succeeds then the same happens for production.
  • When a pull request is opened a series of checks are made, against both staging and production:
    • Python Black (Formats python code correctly)
    • Python iSort (Orders imports correctly)
    • TFLint (Terraform Linter)
    • Terraform Validate
    • Terraform init.
    • Terraform Plan (A plan of the infrastructure changes for that environment)
  • If the checks fail at the pre-commit stage you can usually fix these with pre-commit run --all-files and committing the changes. Problems which can't be auto-fixed will be explained.
  • TFLint will explain any errors it finds.
  • Terraform plan needs to be inspected before merging code to ensure the right thing is being applied. Do not assume that a green build is going to build what you want to be built.
  • Upon merge, staging environment docker images will be built and pushed to ECR, staging environment Terraform code will be applied. On success of the staging environment, production environment docker images will be built and pushed to ECR, production environment Terraform code will be applied.

7 DB Backups

As we use AWS Aurora, there is no multi-AZ functionality. Instead, “Aurora automatically replicates storage six ways across three availability zones”.

Each night there is an automated snapshot by Amazon of RDS. We also run a manual snapshot of the cluster at midday (UTC) each day. This is cron-based from Amazon Eventbridge that triggers a lambda. DB backups are shown in the RDS console under manual snapshots.

8 Terraform Infrastructure

Here are some brief notes on extending the infrastructure.

  • The file terraform/main.tf will invoke each of the modules, more of the same services can be created by adding to those modules. If more modules are created then terraform/main.tf will need to be extended to invoke them.
  • Adding an S3 bucket is done by invoking the secure_bucket module, located at terraform/modules/secure_bucket/, you can see how the existing buckets are created by viewing terraform/modules/lambda_s3/bucket.tf, new buckets should be created by adding to this file. If a bucket policy is added, then an extra statement will automatically be added that denies insecure transport.
  • Docker images are stored in ECR. Each repo needs to exist before a docker image can be pushed to ECR. These are created in terraform/modules/lambda_s3/lambda.tf.

You can find auto-generated documentation on the Terraform resources in terraform/README.md.

Updating the .terraform.lock.hcl file

  • Run terraform init --upgrade -backend=false locally

9 Turning Enrichment Off

There are a number of places where enrichment can be turned off:

  • Marklogic Username/Password

    • Go to the production Marklogic interface, "Admin" (top), "Security" (left), "Users" (left), "enrichment-engine" (mid).
    • Make sure you know where the password is stored so you can put access back afterwards!
    • Changing the password will mean no Enrichment processes can interact with Marklogic -- no getting documents, no uploading them
    • Messages will still be sent from the Editor interface, and will build up.
    • Purge the queues in AWS before turning Enrichment back on, unless you're confident there is nothing bad in there.
    • Seemed to work well last time, but there were a lot of warnings
  • AWS Lambda that fetches XML

    • Not actually tested in anger!
    • Log into da-caselaw-enrichment. Make sure to switch to eu-west-2/London.
    • In Lambda, Functions, select tna-s3-tna-production-fetch-xml
    • Top left, press Throttle.
    • This will prevent any ingestion of the incoming messages which will build up
    • Anything currently in process will finish and will continue to run to completion
    • You can change the concurrency settings to unthrottle it
    • Note that manual changes to the lambda settings will likely be lost if new code is deployed
  • Modifying the code

    • We could change either the code in one of the lambdas -- probably fetch_xml or push_enriched_xml for the start/end of the process
    • Not entirely confident that the lambdas are being automatically deployed correctly at this time
    • We could also modify the privileged API, but that potentially affect all users of it but there aren't any at the time of writing

10 Debugging

Situations when you may want to debug an enrichment run:

  • a document we expect to have been enriched, has not been enriched as expected.
  • an AWS error alert for a lambda function has been raised to us (probably through email notification subscription)

The main ways we have to debug are to look at AWS logs for lambda functions and data stored in s3 buckets but we need appropriate information to find these in AWS first.

You will need access to for the staging or production Enrichment AWS space as appropriate to follow these debugging tips. If not, you could skip most of this and attempt at recreating a local test of a lambda function you think might have a problem like in Recreate and debug

Getting information to investigate in AWS

Find judgment name with failed lambda time

  1. Look at the logs for that lambda function in AWS and look for the logs from around the time of the error alert.
  2. Find the name of the judgment from these log0s

Find time from judgment name

  1. Search for the judgment name in one of the enrichment s3 buckets (probably best to start with the first bucket in case there was a failure and so never reached the later buckets) to see when the bucket was updated.

Inspecting logs and s3 buckets in AWS

Look at lambda logs

Each lambda function has a log group in which we can access different logs for different runs of the lambda.

  1. We can find the logs for a particular lambda at a particular time by filtering the logs in that lambda's log group around the time we know the lambda run we care about was triggered.
  2. Then we can find any relevant information from the logs to use for

Look at s3 buckets

At each stage of enrichment process we store the partially enriched xml back to an s3 bucket as shown in docs/img/architecture.png. You can find all the s3 bucket names where they are defined in terraform/modules/lambda_s3/bucket.tf and can find all relationships between buckets and lambdas in terraform/modules/lambda_s3/lambda.tf.

  1. We can download the xml from each stage for the judgment by going to each s3 bucket in the AWS Enrichment space and searching for the judgment name.
  2. We can compare sequential stages of enriched xml for the judgment to attempt to determine where our issue may have arisen.
    • e.g. if we can see that between enrichment stage 1 completing and enrichment stage 2 completing something odd happened to the xml we can focus on the lambda function that is called in between, here the oblique_reference_replacer lambda.

Recreate and debug

Once we have determined the lambda that caused the issue, xml from the s3 buckets on either side of the lambda, and any relevant log information as a starting point, we can use the xml from each as input and expected output fixtures for an end to end test to help us debug locally with breakpoints and fix the bug.

11 Deploy

Currently, the main branch is deployed to staging, and if that doesn't fail, it is then deployed to production.

Release Process

The version should be an integer string (like "1"): note, however, that pre-December 2023 versions were version "0.1.0".

As a part of each pull request that isn't just keeping versions up to date:

  • Update the version number in enrichment_version.string in src/lambdas/determine_legislation_provisions/index.py
  • Update CHANGELOG.md with a brief description of the change
  • Create a release on Github with a tag like v1. This does nothing, but is useful to help us keep track.