Skip to content

methods.md

Latest commit

 

History

History
137 lines (95 loc) · 3.21 KB

methods.md

File metadata and controls

137 lines (95 loc) · 3.21 KB

Methods API Reference

All methods inherit from BaseReformulator and provide the same interface.

Common Interface

All methods support:

# Single query reformulation
result = method.reformulate(query, contexts=None)

# Batch reformulation
results = method.reformulate_batch(queries, contexts=None)

Available Methods

GenQR

Generic keyword expansion using LLM.

reformulator = qg.create_reformulator("genqr", model="gpt-4")

Requires Context: No

GenQR Ensemble

Ensemble of multiple keyword expansion prompts.

reformulator = qg.create_reformulator(
    "genqr_ensemble",
    model="gpt-4",
    params={"repeat_query_weight": 3}
)

Requires Context: No
Parameters:

  • repeat_query_weight (int): Number of query repetitions (default: 3)

Query2Doc

Generates pseudo-documents for the query.

reformulator = qg.create_reformulator("query2doc", model="gpt-4")

Requires Context: No

QA Expand

Question-answer based expansion.

reformulator = qg.create_reformulator("qa_expand", model="gpt-4")

Requires Context: No

MuGI

Multi-granularity information expansion.

reformulator = qg.create_reformulator("mugi", model="gpt-4")

Requires Context: No

LameR

Context-based passage synthesis.

reformulator = qg.create_reformulator("lamer", model="gpt-4")

Requires Context: Yes

Query2E

Query to entity expansion.

reformulator = qg.create_reformulator("query2e", model="gpt-4")

Requires Context: No

CSQE

Context-based sentence extraction.

reformulator = qg.create_reformulator("csqe", model="gpt-4")

Requires Context: Yes

ThinkQE

Multi-round query expansion with retrieved passage feedback. Each round uses the original query plus newly retrieved passages to generate pseudo-passages via a reasoning LLM, appends them to the retrieval query, and retrieves again.

reformulator = qg.create_reformulator(
    "thinkqe",
    model="deepseek-ai/DeepSeek-R1-Distill-Qwen-14B",
    params={
        "searcher": searcher,
        "num_interaction": 3,
        "keep_passage_num": 5,
        "gen_num": 2,
        "accumulate": True,
        "use_passage_filter": True,
        "search_k": 1000,
    },
    llm_config={"temperature": 0.7, "max_tokens": 32768}
)

Requires Context: Yes
Parameters:

  • keep_passage_num (int): Passages kept for prompting per round (default: 5)
  • gen_num (int): Expansions generated per round (default: 2)
  • num_interaction (int): Expansion rounds after baseline retrieval (default: 3)
  • accumulate (bool): Accumulate all round expansions into later rounds (default: True)
  • use_passage_filter (bool): Blacklist passages repeated from two rounds ago (default: True)
  • repeat_weight (float): Divisor for adaptive query repetition heuristic (default: 3)
  • search_k (int): Retrieval depth per round; use 1000 to mirror original paper runs (default: keep_passage_num)
  • max_demo_len (int): Optional word truncation for each passage before prompting (default: None)
  • no_thinking (bool): Prefill </think> to disable reasoning traces (default: False)
  • searcher: Pre-configured searcher instance (recommended)