Skip to content

Generators

PromptedCitedGenerator

PromptedCitedGenerator 

PromptedCitedGenerator(model: str = 'claude-haiku-4-5-20251001', temperature: float = 0.0, max_tokens: int = 1024, system_prompt: str | None = None, api_base: str | None = None, num_retries: int = 3)

Baseline cited generator using prompt instructions + LiteLLM.

Parameters

model: LiteLLM model identifier, e.g. "claude-haiku-4-5-20251001" or "openai/gpt-4o-mini". See LiteLLM docs for the full list. temperature: Sampling temperature. Default 0 for deterministic citations. max_tokens: Hard cap on generated tokens. Defaults to 1024. system_prompt: Override the default system prompt. api_base: Override LiteLLM's default API base (e.g. for local models).

Source code in src/verifiable_rag/generators/prompted.py 
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
def __init__(
    self,
    model: str = "claude-haiku-4-5-20251001",
    temperature: float = 0.0,
    max_tokens: int = 1024,
    system_prompt: str | None = None,
    api_base: str | None = None,
    num_retries: int = 3,
) -> None:
    if not (0.0 <= temperature <= 2.0):
        raise ValueError(f"temperature must be in [0, 2], got {temperature}")
    if max_tokens < 1:
        raise ValueError(f"max_tokens must be positive, got {max_tokens}")
    if num_retries < 0:
        raise ValueError(f"num_retries must be >= 0, got {num_retries}")

    self._model = model
    self._temperature = temperature
    self._max_tokens = max_tokens
    self._system_prompt = system_prompt or _DEFAULT_SYSTEM_PROMPT
    self._api_base = api_base
    self._num_retries = num_retries

generate 

generate(query: str, chunks: list[RetrievedChunk], documents: dict[str, Document]) -> list[CitedSentence]

Call the LLM and parse its cited output.

Source code in src/verifiable_rag/generators/prompted.py 
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
def generate(
    self,
    query: str,
    chunks: list[RetrievedChunk],
    documents: dict[str, Document],
) -> list[CitedSentence]:
    """Call the LLM and parse its cited output."""
    if not chunks:
        return []

    source_lines, valid_ids = self._format_sources(chunks, documents)
    if not source_lines:
        # No retrievable sentence text — generator cannot ground anything.
        return []

    user_prompt = _USER_PROMPT_TEMPLATE.format(
        query=query.strip(),
        sources="\n".join(source_lines),
    )

    raw = self._call_llm(user_prompt)
    if self._looks_like_refusal(raw):
        return []

    return self._parse_output(raw, valid_ids)

ConstrainedCitedGenerator

ConstrainedCitedGenerator 

ConstrainedCitedGenerator(model: str = 'anthropic/claude-haiku-4-5', temperature: float = 0.0, max_tokens: int = 2048, system_prompt: str | None = None, api_base: str | None = None, num_retries: int = 3, max_citations_per_sentence: int = _DEFAULT_MAX_CITATIONS)

Constrained-decoding cited generator using LiteLLM structured outputs.

Parameters

model: LiteLLM model identifier. Must support response_format json_schema (Anthropic Sonnet/Haiku, OpenAI GPT-4o family, etc.). temperature: Sampling temperature. Default 0 for deterministic citations. max_tokens: Hard cap on generated tokens. Defaults to 2048 (structured JSON is more verbose than free prose). system_prompt: Override the default system prompt. api_base: Override LiteLLM's default API base (e.g. for local models). num_retries: LiteLLM-level retry count for transient errors. max_citations_per_sentence: Schema-enforced cap on the citations array per output sentence. Default 3.

Source code in src/verifiable_rag/generators/constrained.py 
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
def __init__(
    self,
    model: str = "anthropic/claude-haiku-4-5",
    temperature: float = 0.0,
    max_tokens: int = 2048,
    system_prompt: str | None = None,
    api_base: str | None = None,
    num_retries: int = 3,
    max_citations_per_sentence: int = _DEFAULT_MAX_CITATIONS,
) -> None:
    if not (0.0 <= temperature <= 2.0):
        raise ValueError(f"temperature must be in [0, 2], got {temperature}")
    if max_tokens < 1:
        raise ValueError(f"max_tokens must be positive, got {max_tokens}")
    if num_retries < 0:
        raise ValueError(f"num_retries must be >= 0, got {num_retries}")
    if max_citations_per_sentence < 1:
        raise ValueError(
            f"max_citations_per_sentence must be >= 1, got {max_citations_per_sentence}"
        )

    self._model = model
    self._temperature = temperature
    self._max_tokens = max_tokens
    self._system_prompt = system_prompt or _DEFAULT_SYSTEM_PROMPT
    self._api_base = api_base
    self._num_retries = num_retries
    self._max_citations = max_citations_per_sentence

SAFECitedGenerator

SAFECitedGenerator 

SAFECitedGenerator(model: str = 'anthropic/claude-haiku-4-5', temperature: float = 0.0, max_tokens: int = 4096, system_prompt: str | None = None, api_base: str | None = None, num_retries: int = 3, max_citations_per_claim: int = _DEFAULT_MAX_CITATIONS_PER_CLAIM, max_claims_per_sentence: int = _DEFAULT_MAX_CLAIMS_PER_SENTENCE)

SAFE-style atomic-claim cited generator using LiteLLM structured outputs.

Functionally equivalent to :class:ConstrainedCitedGenerator with one architectural addition: each sentence is decomposed into atomic claims, and citations attach to atomic claims rather than whole sentences. This enables fine-grained verification on benchmarks with atomic-claim-level gold (RAGTruth, FaithBench).

For benchmarks with sentence-level gold (LitQA2, ALCE), the parser flattens atomic_claims back into our existing CitedSentence shape so existing eval metrics work unchanged.

Parameters

model: LiteLLM model identifier. Must support response_format json_schema (Anthropic Sonnet/Haiku, OpenAI GPT-4o family, etc.). temperature: Sampling temperature. Default 0 for deterministic citations. max_tokens: Hard cap on generated tokens. Defaults to 4096 — atomic decomposition is more verbose than flat constrained output. system_prompt: Override the default system prompt. api_base: Override LiteLLM's default API base (e.g. for local models). num_retries: LiteLLM-level retry count for transient errors. max_citations_per_claim: Schema-enforced cap on citations per atomic claim. Default 3. max_claims_per_sentence: Schema-enforced cap on atomic claims per output sentence. Default 5.

Source code in src/verifiable_rag/generators/safe.py 
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
def __init__(
    self,
    model: str = "anthropic/claude-haiku-4-5",
    temperature: float = 0.0,
    max_tokens: int = 4096,
    system_prompt: str | None = None,
    api_base: str | None = None,
    num_retries: int = 3,
    max_citations_per_claim: int = _DEFAULT_MAX_CITATIONS_PER_CLAIM,
    max_claims_per_sentence: int = _DEFAULT_MAX_CLAIMS_PER_SENTENCE,
) -> None:
    if not (0.0 <= temperature <= 2.0):
        raise ValueError(f"temperature must be in [0, 2], got {temperature}")
    if max_tokens < 1:
        raise ValueError(f"max_tokens must be positive, got {max_tokens}")
    if num_retries < 0:
        raise ValueError(f"num_retries must be >= 0, got {num_retries}")
    if max_citations_per_claim < 1:
        raise ValueError(
            f"max_citations_per_claim must be >= 1, got {max_citations_per_claim}"
        )
    if max_claims_per_sentence < 1:
        raise ValueError(
            f"max_claims_per_sentence must be >= 1, got {max_claims_per_sentence}"
        )

    self._model = model
    self._temperature = temperature
    self._max_tokens = max_tokens
    self._system_prompt = system_prompt or _DEFAULT_SYSTEM_PROMPT
    self._api_base = api_base
    self._num_retries = num_retries
    self._max_cites = max_citations_per_claim
    self._max_claims = max_claims_per_sentence