docs(examples): add heterogeneous financial document RAG evaluation example with threshold_overrides

[Ruthwik-Data]

Summary

Adds examples/rag_evaluation/heterogeneous_document_eval.py — a runnable example demonstrating how to evaluate a financial RAG pipeline that retrieves chunks from mixed document types (balance sheets, 10-K narratives, earnings call transcripts) using document-type-specific thresholds.

Relates to #2775 and complements PR #2785 (threshold_overrides implementation).

Why this example is needed

Issue #2775 raised a real gap: the existing examples/rag_evaluation/ directory has no example covering heterogeneous corpora. A single global threshold causes systematic false failures or false passes when evaluating mixed document types:

• Balance sheets (structured, numeric): threshold 0.95 — hallucinated figures are dangerous
• 10-K narratives (dense text): threshold 0.80 — some paraphrase acceptable
• Earnings calls (hedged, forward-looking): threshold 0.70 — hedged language should not fail faithfulness

This example shows users exactly how to use additional_metadata["document_type"] today (readable pattern), with commented-out threshold_overrides lines that will activate once PR #2785 is merged.

What this PR adds

examples/rag_evaluation/heterogeneous_document_eval.py

Sections:

1. THRESHOLD_OVERRIDES — dict mapping document types to thresholds (balance_sheet: 0.95, annual_report: 0.80, earnings_call: 0.70)
2. build_test_cases() — three LLMTestCase instances with document_type metadata for each document type
3. build_metrics() — FaithfulnessMetric, ContextualPrecisionMetric, ContextualRecallMetric with commented threshold_overrides lines (forward-compatible with PR feat(metrics): add document-type threshold overrides for heterogeneous RAG evaluation #2785)
4. run_heterogeneous_eval() — runs evaluate() and prints per-document-type scores with expected thresholds

Design notes

• Works today without PR feat(metrics): add document-type threshold overrides for heterogeneous RAG evaluation #2785 (uses default threshold); threshold routing is forward-compatible via # Uncomment post-PR #2785 comments
• Uses public financial data patterns (no real company data)
• Requires only OPENAI_API_KEY — no additional dependencies
• Follows existing pattern in examples/rag_evaluation/

Type of change

• Documentation (non-breaking addition or improvement to examples/docs)

[vercel]

@Ruthwik-Data is attempting to deploy a commit to the Confident AI Team on Vercel.

A member of the Team first needs to authorize it.