AI Navigation Map

Operation

Map the requested area using current repository evidence.

Goal

Map the selected AI capabilities from input and model calls through prompts, tools or retrieval to outputs and evaluation.

Suggested saved map: docs/ai_map.md.

Build the view

Trace one actual feature entry through input preparation, retrieval/tool selection, model invocation, parsing and the consuming surface. Follow the prompt and schema owners rather than grouping everything under the SDK wrapper. Connect each claimed quality/safety property to the evaluation or enforcement path that observes it; an evaluation file's existence does not establish a passing result. Expand to other requested capabilities using this same trace, with model/index versions and unverified edges explicit.

Inspect

Model invocation sites, SDK clients, prompt templates, system/developer/user message builders, tool/function schemas, structured-output parsers, streaming handlers, retry/cache layers, rate-limit code, queues/jobs, RAG ingestion/retrieval, vector indexes, OCR/document processing, ML model artifacts, feature stores, eval scripts, fixtures, tests, telemetry/logging, config/env validation, docs, and generated artifacts.

Map content

Evidence and IDs

• Suggested IDs for new entries: capability AI-*, model call MC-*, prompt/template PT-*, tool/function TF-*, retriever/index RG-*, OCR/doc pipeline DOC-*, ML model/pipeline ML-*, eval EV-*, safety boundary SB-*.

Coverage

Cover user-visible AI features, internal/background AI jobs, model/provider configuration, input/output contracts, tool/function contracts, schema validators, streaming and async state, RAG/OCR/ML sub-pipelines, eval coverage, data/privacy boundaries, cost/rate-limit controls, logs/metrics/traces, fallbacks, and operator runbooks when discoverable.

Entry details

For each AI capability, capture owner symbol, caller surface, input sources, model/provider/config, prompt/template owner, output schema, downstream consumers, validation/parsing, error/fallback behavior, privacy/security boundary, tests/evals, telemetry, cost/rate-limit controls, and known drift. For RAG/OCR/ML entries, include ingestion/source data, transformation steps, index/model artifact owners, freshness/versioning, and acceptance/eval signals.