Supported

Knowledge parsers¶

Behavior¶

Mari provides	Caller provides
Typed schemas, evidence resolution, ordering recovery, and safe fallbacks	Model prompt, inference, and task-specific quality
An uncertain result when a required row or citation is invalid	The threshold for retry, review, or rejection

The parsers validate injected extraction results. Corpus evidence helps select the model. Parser guarantees cover structure, attribution, and failure behavior.

Example malformed-batch recovery

Requested claim	Model row	Parsed result
Claim A	Valid supported row with evidence	Supported assessment
Claim B	Missing row	Uncertain assessment
Claim C	Malformed evidence	Uncertain assessment with parse issue
Claim D	Valid row returned out of order	Restored to Claim D position

Models return JSON-like values. Parsers resolve evidence against supplied document and section revisions and return immutable typed values. Research establishes each task formulation. Mari implements the deterministic validation boundary around the cited model.

How it works¶

Each parser requires the recipe’s top-level collection, validates every required field and enum, resolves evidence through the contract below, derives audit signals, and constructs frozen result types. Claim meaning stays with the caller. Batch claim assessment keys rows to caller order. Absent rows become uncertain, and malformed rows leave valid siblings available.

model proposal schema exact evidence revision binding typed candidate

Parser	Produces	Research-backed task	Academic sources
`parse_facts`	`FactCandidate`	Atomic claims and optional document-level relations with evidence	FActScore · DocRED
`parse_claim_assessments`	`FactAssessment`	Supported, contradicted, or uncertain verdicts. Decisive rows require evidence	FEVER
`parse_decisions`	`DecisionCandidate`	Decision-related utterance extraction with evidence-based interpretation	Hsueh & Moore · Karan et al.
`parse_answer`	`GroundedAnswer`	Evidence-selected document QA, citations, or explicit insufficient evidence	QASPER · ALCE
`parse_answer_candidates`	`AnswerCandidate[]`	Reusable question-answer pairs bound to supporting passages	QASPER
`parse_glossary`	`GlossaryCandidate[]`	Term-definition relations, aliases, and source spans	DeftEval
`parse_digest`	`DigestSummary`	Document-wide and topic summaries with separately inspectable evidence	QAGS · SummaC
`parse_impact`	`ImpactAssessment`	In-scope affected-document proposals followed by deterministic dependency checks	Mari contract. No claimed benchmark reproduction
`parse_refinement`	`RefinementEdit[]`	Bounded, attribution-aware, fact-preserving edit proposals	RARR · FactEditor

answer.py¶

from mari_kit.knowledge import parse_answer

raw = model(question, documents)
answer = parse_answer(question, documents, raw)
print(answer.disposition)         # grounded | insufficient_evidence
print(answer.grounding_coverage)  # deterministic text coverage
for evidence in answer.evidence:
    print(evidence.quote)  # insufficient-evidence answers can have no citations

Additional deterministic helpers include normalize_claim, deduplicate_fact_candidates, grounding_coverage, and excerpt. Recoverable batch drift is handled conservatively: assessment rows are restored to caller order, missing rows become uncertain, and good rows survive alongside invalid ones. Structured fact qualifiers preserve subject, relation, object, scope, validity, and conditions.

grounding_coverage measures lexical coverage. It is a Mari-specific audit signal motivated by citation completeness. Exact quote validation catches fabricated citations. Logical entailment remains a separate evaluation.

Source parser functions and options¶

Function	Main options	Failure behavior
`parse_markdown`	`tables`, `recover_unclosed_fences`, `parser_id`	Returns an unclosed-fence warning or error and retains the recovered block
`parse_html`	`parser_id`, `recover`	Reports unmatched and implicitly closed tags. `recover=False` promotes recovery warnings to errors
`parse_delimited`	`identity_fields`, delimiter or sniffing, quote character, header, `strict_width`	Rejects a malformed-width row in strict mode. Valid siblings survive
`parse_json_lines`	`identity_fields`, `parser_id`	Each malformed or non-object line is positioned independently
`parse_json_array`	`identity_fields`, `parser_id`	Retains parsed object siblings and positions non-object or malformed members
`parse_python`	Repository, revision, path and parser ID	Syntax errors produce an empty, positioned result. Symbol extraction requires valid syntax

ParseIssue contains code and message. Severity and an optional subject add context. The record also carries an exact character span. ParseResult.succeeded becomes true when every issue stays below error severity. Warnings and accepted values remain available in the result.

Keep accepted records beside malformed siblings¶

from mari_kit.documents import parse_json_lines

result = parse_json_lines(
    payload,
    source_id="warehouse:policies",
    revision=object_etag,
    identity_fields=("policy_id", "jurisdiction"),
)

store(result.values)
observe(result.issues)

Structured-ingest fixture	Before	Current
Fields with exact raw spans	`0/12`	`12/12`
Syntactically accepted records with record spans	`0/5`	`5/5`
Schema-drift violations	`1`	`3`
Invalid integer/date values rejected	`0/2`	`2/2`

Format and parser references

CommonMark block syntax WHATWG HTML parsing RFC 4180 CSV JSON Lines Python AST Tree-sitter

Markdown-It-Py and html5lib are MIT-licensed differential references. Tree-sitter and its Python grammar are MIT licensed.