Rank prefix and next-word candidates · AI Engineering

Type: Build Language: TypeScript Stage: 2 of 4 Time: ~2 hours Prerequisites: Basic TypeScript, JSON objects, arrays and running Node 22.18 or newer. Complete the preceding stages first.

What you build

Rank candidates by the longest observed history and show the underlying count.

Public contract

typescript

suggest(model: Model, prefix: string, history?: string[], limit?: number): Suggestion[]

Normalize the prefix; filter words by startsWith. For each word use the longest matching suffix history (2 then 1), otherwise its unigram count. Sort by order descending, count descending, then lexical word. Return {word,score,order,count,sources}; score is order * 1000000 + count for display, sorting uses separate order/count. limit defaults to 5 and must be an integer 0..50.

The initial workspace contains public types, function stubs, the CLI, and original input fixtures. Change main.ts; the CLI is already supplied. Stages accumulate without replacing your earlier implementation. Tests import the learner workspace selected by PROJECT_WORKSPACE.

Worked example

For prefix s after history ["garden","club"], shares with order 2/count 1 precedes seed with order 0/count 20. A low raw count can still be stronger conditional evidence.

pj vocabulary autocomplete pad 2

Change the figure controls, predict the intermediate values, and compare the calculation with your implementation. The figure explains the mechanism; the grader runs the real functions.

Implement and verify

bash

python3 scripts/project_test.py vocabulary-autocomplete-pad --init my-vocabulary-autocomplete-pad
python3 scripts/project_test.py vocabulary-autocomplete-pad --stage 2 --path my-vocabulary-autocomplete-pad --strict

Initialize once. The fresh first stage intentionally fails with a named not-implemented error. Preserve the signatures and reject malformed inputs with useful errors. Keep inputs unchanged. Each stage includes separate held-out inputs; passing the demonstration alone is insufficient.

Failure cases and hints

Unseen history backs off. A prefix with no candidates returns []; no remote completion is fabricated.

Use your result

After stage four, pass a JSON file of your own through the same entry point:

bash

cd my-vocabulary-autocomplete-pad
node cli.ts sample.json output

A local autocomplete writing page and vocabulary-model.json. Open output/report.html and inspect the machine-readable companion files.

Extend it

Add smoothing while preserving raw evidence counts.

Scope: This is a transparent count-based local language model, not a hosted LLM. It suggests a single next token with up to two tokens of context. The browser editor operates at the end of input. No provider adapter is implied.