Compiler profiles

A compiler profile is the primary configuration surface for controlling what riverbank extracts, how it evaluates quality, and what the compiled graph must be able to answer.

What a profile controls

Field	Purpose
name	Unique identifier for the profile
version	Integer version (supports multiple versions per name)
extractor	Which extractor plugin to use (noop, instructor, or custom)
model_provider	LLM provider (ollama, openai, anthropic, vllm, azure-openai)
model_name	Model identifier
embed_model	Embedding model for vector operations
max_fragment_tokens	Maximum tokens per fragment sent to the LLM
named_graph	Target named graph IRI for compiled triples
prompt_text	System prompt guiding extraction
editorial_policy	Gate rules for fragment filtering
competency_questions	SPARQL assertions the compiled graph must satisfy
run_mode_sequence	Multi-pass compilation order (e.g., [vocabulary, full])
absence_rules	Rules for recording explicit absences
ensemble	Multi-model configuration for higher accuracy

Profile as a contract

The profile defines a contract between the corpus author and the knowledge consumer:

1. What gets extracted — controlled by prompt_text and extractor
2. What quality level is acceptable — controlled by editorial_policy.confidence_threshold
3. What questions the graph must answer — controlled by competency_questions

This contract is testable. The golden corpus CI gate validates competency questions on every commit.

Competency-question-driven design

The most effective way to design a profile:

1. Write the SPARQL queries your consumers need to answer
2. Put them in competency_questions
3. Iterate on prompt_text until all questions pass
4. The competency questions become your regression test suite

Profile versioning

Profiles are versioned by (name, version). When you improve a profile, bump the version. Old versions remain in the catalog for audit trail and recompilation comparison.

# Register v2 of an existing profile
riverbank profile register improved-profile-v2.yaml

Editorial policy

The editorial policy prevents wasted LLM calls on content that cannot produce useful knowledge:

editorial_policy:
  min_fragment_length: 50      # Characters — skip tiny fragments
  max_fragment_length: 8000    # Characters — flag oversized fragments
  min_heading_depth: 0         # 0 = all headings; 2 = skip H1
  confidence_threshold: 0.7    # Below this → draft graph
  allowed_languages: [en]      # Skip non-English content

See the compiler profile schema reference for exhaustive field documentation.