Knowledge

Scrivai provides three first-class library types for injecting domain knowledge into PES runs. All three are backed by the qmd semantic retrieval engine.

Library Types

Type	Class	Purpose
Rules	RuleLibrary	Audit criteria, constraints, and compliance requirements
Cases	CaseLibrary	Reference examples (few-shot demonstrations, precedents)
Templates	TemplateLibrary	Document templates with {{variable}} placeholders

Setup via build_libraries

The recommended way to initialise all three libraries at once is build_libraries, which reads collection paths from a config dict:

from scrivai import build_libraries, build_qmd_client_from_config

# Build the qmd client from project config
qmd_client = build_qmd_client_from_config(
    config={"embedding_model": "text-embedding-3-small"}
)

# Build all three libraries
rule_lib, case_lib, template_lib = build_libraries(
    qmd_client=qmd_client,
    rules_path="knowledge/rules/",
    cases_path="knowledge/cases/",
    templates_path="knowledge/templates/",
)

Search Interface

Each library exposes a .search(query, top_k) method that returns semantically ranked results:

# Find the most relevant audit rules for a query
results = rule_lib.search("financial disclosure requirements", top_k=5)
for result in results:
    print(result.chunk_text)
    print(result.score)

# Find reference cases
cases = case_lib.search("contract termination clause", top_k=3)

# Find a matching template
templates = template_lib.search("engineering inspection report", top_k=1)

Injecting Knowledge into a PES

Pass search results into runtime_context so the PES can use them in its system prompt:

rules = rule_lib.search("safety compliance", top_k=5)
result = pes.run(
    runtime_context={
        "document_text": document,
        "applicable_rules": [r.chunk_text for r in rules],
    }
)

See Also

• API Reference: Knowledge