Skip to content

lpgcast.repair

The typed repair prompt: the findings of one reply, grouped by kind.

Each kind gets one instruction and the fragments it concerns, nothing else. The model is asked for the whole reply again, corrected, so nothing has to be merged back. Consequences of other findings (an edge left out because its vertex was) are not listed; they clear when the cause does.

Attributes

DETAIL_GUIDANCE = {('coercion', 'unit_mismatch'): 'A property shown with a unit in [brackets] takes its value in exactly that unit, and the symbol is case-sensitive (mW is not MW). If the text states the value in another unit, leave the property out; never convert it.'} module-attribute

GUIDANCE = {'unknown_type': 'Use only types listed under TYPES. If no type fits, move the item to `unmapped`.', 'duplicate_ref': 'Give every vertex its own ref.', 'identity_insufficient': 'Give the key properties (`id:`) the text states. If the text does not state them, leave the vertex out or move it to `unmapped`; never invent a value.', 'undeclared_edge': 'Use only the edge patterns listed under TYPES, with `from` of the source type and `to` of the target type. If none fits, move the relation to `unmapped`.', 'dangling_ref': '`from` and `to` must be refs of vertices in your reply.', 'undeclared_property': 'Use only the properties listed for that type.', 'duplicate_property': 'Give each property at most once per vertex or edge.', 'coercion': "Write each value so it fits the property's type: a plain number for INT and FLOAT (1250, not 1,250), true or false for BOOL, an ISO 8601 date for DATETIME (2024-03-04, or 2024-03 or 2024 when the text gives no more). Leave the property out if the text does not state it.", 'value_not_in_quote': 'The value must be the one its `quote` states: the same digits, and the same day, month and year for a date. Correct the value or the quote; if the text does not state the value, leave the property out.', 'quote_not_found': 'Every `quote` must be copied character for character from the TEXT. Do not paraphrase, shorten with ellipses or quote CONTEXT, except for a key of an entity the TEXT refers to.', 'endpoint_invalid': ''} module-attribute

INTRO = '# CORRECTIONS\n\nYour reply has the problems listed below, grouped by kind. Reply again with the whole extraction in the same format, with these fixed and everything else kept as it was. Remove an item, or move it to `unmapped`, when the text does not support it.' module-attribute

Classes

Functions:

repair_messages(messages, previous_reply, findings)

The original request, the reply that failed, and the corrections.

Source code in lpgcast/repair.py 
 93
 94
 95
 96
 97
 98
 99
100
101
def repair_messages(
    messages: list[ChatMessage], previous_reply: str, findings: list[Finding]
) -> list[ChatMessage]:
    """The original request, the reply that failed, and the corrections."""
    return [
        *messages,
        {"role": "assistant", "content": previous_reply},
        {"role": "user", "content": repair_prompt(findings)},
    ]

repair_prompt(findings)

The corrections message for findings. Deterministic for a given list.

Source code in lpgcast/repair.py 
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
def repair_prompt(findings: list[Finding]) -> str:
    """The corrections message for *findings*. Deterministic for a given list."""
    by_kind: dict[FindingKind, list[Finding]] = {}
    for finding in findings:
        by_kind.setdefault(finding.kind, []).append(finding)
    sections = [INTRO]
    for kind in FINDING_ORDER:
        group = by_kind.get(kind)
        if not group or not GUIDANCE[kind]:
            continue
        lines = [f"## {kind}", "", GUIDANCE[kind], ""]
        details = sorted({f.detail for f in group if f.detail is not None})
        extra = [
            DETAIL_GUIDANCE[(kind, d)] for d in details if (kind, d) in DETAIL_GUIDANCE
        ]
        if extra:
            lines[-1:-1] = ["", *extra]
        for finding in group:
            fragment = json.dumps(
                finding.fragment, ensure_ascii=False, separators=(",", ":")
            )
            lines.append(f"- `{finding.path}`: {finding.message}\n  {fragment}")
        sections.append("\n".join(lines))
    return "\n\n".join(sections) + "\n"