Skip to content

Extractors

Built-in parsing helpers are registered automatically by PluginRegistry.

builtin

Built-in extractor implementations auto-registered by PluginRegistry.

BoxedTextExtractor

Extract the final LaTeX-style boxed answer from raw text.

Source code in themis/extractors/builtin.py 
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
class BoxedTextExtractor:
    """Extract the final LaTeX-style boxed answer from raw text."""

    def extract(
        self,
        trial: TrialSpec,
        candidate: CandidateRecord,
        config: Mapping[str, JSONValueType] | None = None,
    ) -> ExtractionRecord:
        """Extract the final boxed segment from candidate output."""
        del trial
        cfg = dict(config or {})
        boxed = _last_boxed_text(_raw_text(candidate))
        if boxed is None:
            return _failure(
                "boxed_text",
                candidate,
                cfg,
                "No boxed answer found in the inference output.",
            )
        return _success("boxed_text", candidate, cfg, boxed)

extract 

extract(
    trial: TrialSpec,
    candidate: CandidateRecord,
    config: Mapping[str, JSONValueType] | None = None,
) -> ExtractionRecord

Extract the final boxed segment from candidate output.

Source code in themis/extractors/builtin.py 
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
def extract(
    self,
    trial: TrialSpec,
    candidate: CandidateRecord,
    config: Mapping[str, JSONValueType] | None = None,
) -> ExtractionRecord:
    """Extract the final boxed segment from candidate output."""
    del trial
    cfg = dict(config or {})
    boxed = _last_boxed_text(_raw_text(candidate))
    if boxed is None:
        return _failure(
            "boxed_text",
            candidate,
            cfg,
            "No boxed answer found in the inference output.",
        )
    return _success("boxed_text", candidate, cfg, boxed)

ChoiceLetterExtractor

Extract an uppercase multiple-choice letter from candidate raw text.

Source code in themis/extractors/builtin.py 
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
class ChoiceLetterExtractor:
    """Extract an uppercase multiple-choice letter from candidate raw text."""

    def extract(
        self,
        trial: TrialSpec,
        candidate: CandidateRecord,
        config: Mapping[str, JSONValueType] | None = None,
    ) -> ExtractionRecord:
        """Extract a multiple-choice answer letter from candidate output."""
        del trial
        cfg = dict(config or {})
        configured_choices = cfg.get("choices")
        if isinstance(configured_choices, list) and configured_choices:
            choices = [str(choice).upper() for choice in configured_choices]
        else:
            choices = ["A", "B", "C", "D", "E"]
        choices_pattern = "".join(re.escape(choice) for choice in choices)
        text = _last_boxed_text(_raw_text(candidate)) or _raw_text(candidate)

        match = re.search(
            rf"\b(?:option|answer|choice)\s*[:\-]?\s*([{choices_pattern}])\b",
            text,
            flags=re.IGNORECASE,
        )
        if match is None:
            match = re.search(rf"\b([{choices_pattern}])\b", text, flags=re.IGNORECASE)
        if match is None:
            return _failure(
                "choice_letter",
                candidate,
                cfg,
                "No choice letter found in the inference output.",
            )
        return _success("choice_letter", candidate, cfg, match.group(1).upper())

extract 

extract(
    trial: TrialSpec,
    candidate: CandidateRecord,
    config: Mapping[str, JSONValueType] | None = None,
) -> ExtractionRecord

Extract a multiple-choice answer letter from candidate output.

Source code in themis/extractors/builtin.py 
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
def extract(
    self,
    trial: TrialSpec,
    candidate: CandidateRecord,
    config: Mapping[str, JSONValueType] | None = None,
) -> ExtractionRecord:
    """Extract a multiple-choice answer letter from candidate output."""
    del trial
    cfg = dict(config or {})
    configured_choices = cfg.get("choices")
    if isinstance(configured_choices, list) and configured_choices:
        choices = [str(choice).upper() for choice in configured_choices]
    else:
        choices = ["A", "B", "C", "D", "E"]
    choices_pattern = "".join(re.escape(choice) for choice in choices)
    text = _last_boxed_text(_raw_text(candidate)) or _raw_text(candidate)

    match = re.search(
        rf"\b(?:option|answer|choice)\s*[:\-]?\s*([{choices_pattern}])\b",
        text,
        flags=re.IGNORECASE,
    )
    if match is None:
        match = re.search(rf"\b([{choices_pattern}])\b", text, flags=re.IGNORECASE)
    if match is None:
        return _failure(
            "choice_letter",
            candidate,
            cfg,
            "No choice letter found in the inference output.",
        )
    return _success("choice_letter", candidate, cfg, match.group(1).upper())

FirstNumberExtractor

Extract the first integer or floating-point token from raw text.

Source code in themis/extractors/builtin.py 
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
class FirstNumberExtractor:
    """Extract the first integer or floating-point token from raw text."""

    def extract(
        self,
        trial: TrialSpec,
        candidate: CandidateRecord,
        config: Mapping[str, JSONValueType] | None = None,
    ) -> ExtractionRecord:
        """Extract the first numeric token from candidate output."""
        del trial
        cfg = dict(config or {})
        text = _raw_text(candidate)
        match = re.search(r"[-+]?\d+(?:\.\d+)?", text)
        if match is None:
            return _failure(
                "first_number",
                candidate,
                cfg,
                "No numeric token found in the inference output.",
            )
        number_text = match.group(0)
        parsed: int | float = (
            int(number_text) if "." not in number_text else float(number_text)
        )
        return _success("first_number", candidate, cfg, parsed)

extract 

extract(
    trial: TrialSpec,
    candidate: CandidateRecord,
    config: Mapping[str, JSONValueType] | None = None,
) -> ExtractionRecord

Extract the first numeric token from candidate output.

Source code in themis/extractors/builtin.py 
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
def extract(
    self,
    trial: TrialSpec,
    candidate: CandidateRecord,
    config: Mapping[str, JSONValueType] | None = None,
) -> ExtractionRecord:
    """Extract the first numeric token from candidate output."""
    del trial
    cfg = dict(config or {})
    text = _raw_text(candidate)
    match = re.search(r"[-+]?\d+(?:\.\d+)?", text)
    if match is None:
        return _failure(
            "first_number",
            candidate,
            cfg,
            "No numeric token found in the inference output.",
        )
    number_text = match.group(0)
    parsed: int | float = (
        int(number_text) if "." not in number_text else float(number_text)
    )
    return _success("first_number", candidate, cfg, parsed)

JsonSchemaExtractor

Parse candidate raw text as JSON and validate it against a schema.

Source code in themis/extractors/builtin.py 
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
class JsonSchemaExtractor:
    """Parse candidate raw text as JSON and validate it against a schema."""

    def extract(
        self,
        trial: TrialSpec,
        candidate: CandidateRecord,
        config: Mapping[str, JSONValueType] | None = None,
    ) -> ExtractionRecord:
        """Parse candidate output as JSON and validate it against a schema."""
        del trial
        cfg = dict(config or {})
        schema = cfg.get("schema")
        if not isinstance(schema, Mapping):
            return _failure(
                "json_schema",
                candidate,
                cfg,
                "Json schema extractor requires a 'schema' object.",
            )

        try:
            parsed = json.loads(_raw_text(candidate))
        except json.JSONDecodeError as exc:
            return _failure(
                "json_schema",
                candidate,
                cfg,
                f"Response was not valid JSON: {exc.msg}.",
            )

        try:
            jsonschema = import_optional("jsonschema", extra="extractors")
        except ThemisError as exc:
            return _failure("json_schema", candidate, cfg, exc.message)

        try:
            validator_cls = jsonschema.validators.validator_for(schema)
            validator_cls.check_schema(schema)
            validator_cls(schema).validate(parsed)
        except jsonschema.exceptions.SchemaError as exc:
            return _failure(
                "json_schema", candidate, cfg, f"Invalid JSON schema: {exc.message}"
            )
        except jsonschema.exceptions.ValidationError as exc:
            return _failure("json_schema", candidate, cfg, exc.message)
        return _success("json_schema", candidate, cfg, parsed)

extract 

extract(
    trial: TrialSpec,
    candidate: CandidateRecord,
    config: Mapping[str, JSONValueType] | None = None,
) -> ExtractionRecord

Parse candidate output as JSON and validate it against a schema.

Source code in themis/extractors/builtin.py 
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
def extract(
    self,
    trial: TrialSpec,
    candidate: CandidateRecord,
    config: Mapping[str, JSONValueType] | None = None,
) -> ExtractionRecord:
    """Parse candidate output as JSON and validate it against a schema."""
    del trial
    cfg = dict(config or {})
    schema = cfg.get("schema")
    if not isinstance(schema, Mapping):
        return _failure(
            "json_schema",
            candidate,
            cfg,
            "Json schema extractor requires a 'schema' object.",
        )

    try:
        parsed = json.loads(_raw_text(candidate))
    except json.JSONDecodeError as exc:
        return _failure(
            "json_schema",
            candidate,
            cfg,
            f"Response was not valid JSON: {exc.msg}.",
        )

    try:
        jsonschema = import_optional("jsonschema", extra="extractors")
    except ThemisError as exc:
        return _failure("json_schema", candidate, cfg, exc.message)

    try:
        validator_cls = jsonschema.validators.validator_for(schema)
        validator_cls.check_schema(schema)
        validator_cls(schema).validate(parsed)
    except jsonschema.exceptions.SchemaError as exc:
        return _failure(
            "json_schema", candidate, cfg, f"Invalid JSON schema: {exc.message}"
        )
    except jsonschema.exceptions.ValidationError as exc:
        return _failure("json_schema", candidate, cfg, exc.message)
    return _success("json_schema", candidate, cfg, parsed)

NormalizedTextExtractor

Normalize free-form text for robust exact-match style scoring.

Source code in themis/extractors/builtin.py 
233
234
235
236
237
238
239
240
241
242
243
244
245
246
class NormalizedTextExtractor:
    """Normalize free-form text for robust exact-match style scoring."""

    def extract(
        self,
        trial: TrialSpec,
        candidate: CandidateRecord,
        config: Mapping[str, JSONValueType] | None = None,
    ) -> ExtractionRecord:
        """Normalize either the boxed answer or the full raw text."""
        del trial
        cfg = dict(config or {})
        source = _last_boxed_text(_raw_text(candidate)) or _raw_text(candidate)
        return _success("normalized_text", candidate, cfg, _normalize_text(source))

extract 

extract(
    trial: TrialSpec,
    candidate: CandidateRecord,
    config: Mapping[str, JSONValueType] | None = None,
) -> ExtractionRecord

Normalize either the boxed answer or the full raw text.

Source code in themis/extractors/builtin.py 
236
237
238
239
240
241
242
243
244
245
246
def extract(
    self,
    trial: TrialSpec,
    candidate: CandidateRecord,
    config: Mapping[str, JSONValueType] | None = None,
) -> ExtractionRecord:
    """Normalize either the boxed answer or the full raw text."""
    del trial
    cfg = dict(config or {})
    source = _last_boxed_text(_raw_text(candidate)) or _raw_text(candidate)
    return _success("normalized_text", candidate, cfg, _normalize_text(source))

RegexExtractor

Extract a regex match or capture group from candidate raw text.

Source code in themis/extractors/builtin.py 
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
class RegexExtractor:
    """Extract a regex match or capture group from candidate raw text."""

    def extract(
        self,
        trial: TrialSpec,
        candidate: CandidateRecord,
        config: Mapping[str, JSONValueType] | None = None,
    ) -> ExtractionRecord:
        """Extract a configured regex match from candidate output."""
        del trial
        cfg = dict(config or {})
        pattern = cfg.get("pattern")
        if not isinstance(pattern, str) or not pattern:
            return _failure(
                "regex",
                candidate,
                cfg,
                "Regex extractor requires a non-empty 'pattern'.",
            )

        group = cfg.get("group", 0)
        text = _raw_text(candidate)
        match = re.search(pattern, text)
        if match is None:
            return _failure(
                "regex", candidate, cfg, "Pattern did not match the inference output."
            )

        if not isinstance(group, (int, str)):
            return _failure(
                "regex",
                candidate,
                cfg,
                "Configured capture group must be an integer or string.",
            )

        try:
            value = match.group(int(group))
        except (IndexError, ValueError, TypeError):
            return _failure(
                "regex",
                candidate,
                cfg,
                "Configured capture group was not present in the regex match.",
            )
        return _success("regex", candidate, cfg, value)

extract 

extract(
    trial: TrialSpec,
    candidate: CandidateRecord,
    config: Mapping[str, JSONValueType] | None = None,
) -> ExtractionRecord

Extract a configured regex match from candidate output.

Source code in themis/extractors/builtin.py 
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
def extract(
    self,
    trial: TrialSpec,
    candidate: CandidateRecord,
    config: Mapping[str, JSONValueType] | None = None,
) -> ExtractionRecord:
    """Extract a configured regex match from candidate output."""
    del trial
    cfg = dict(config or {})
    pattern = cfg.get("pattern")
    if not isinstance(pattern, str) or not pattern:
        return _failure(
            "regex",
            candidate,
            cfg,
            "Regex extractor requires a non-empty 'pattern'.",
        )

    group = cfg.get("group", 0)
    text = _raw_text(candidate)
    match = re.search(pattern, text)
    if match is None:
        return _failure(
            "regex", candidate, cfg, "Pattern did not match the inference output."
        )

    if not isinstance(group, (int, str)):
        return _failure(
            "regex",
            candidate,
            cfg,
            "Configured capture group must be an integer or string.",
        )

    try:
        value = match.group(int(group))
    except (IndexError, ValueError, TypeError):
        return _failure(
            "regex",
            candidate,
            cfg,
            "Configured capture group was not present in the regex match.",
        )
    return _success("regex", candidate, cfg, value)