Skip to content

search_agent_distillation.md

Latest commit

 

History

History
88 lines (71 loc) · 3.69 KB

search_agent_distillation.md

File metadata and controls

88 lines (71 loc) · 3.69 KB
Raw

Search Agent Distillation

The search_agent_distill pipeline synthesizes verified multi-hop search tasks from seed QA pairs and distills search-agent (ReAct) trajectories into standard SFT training data. It ports the SearchSynthAgent closed-loop generation system onto native easydistill operators — no LangGraph runtime dependency; all model calls go through the unified backend abstraction.

Pipeline Overview

Seed QA (id, question, answer)
  │
  ▼
search_task_evolve      Strategist-driven closed loop per seed:
  │                       Strategist → EXPAND / REFINE / ROLLBACK / FINALIZE
  │                       EXPAND  = atomic-QA merge (search target → atomic QA → rewrite)
  │                       REFINE  = FUZZ one clue without adding hops
  │                       QualityGate (uniqueness / pseudo multi-hop / leakage)
  │                       Verify (solver rollout or fast verify) → Judge (difficulty report)
  │                       Finalize gate: solver-correct AND difficulty == "good",
  │                       then N-run final eval (accuracy / avg turns)
  ▼
search_trajectory       repeat_times solver rollouts per evolved task;
  │                     each run judged for answer correctness
  ▼
search_judge_filter     keep correct trajectories, select best (fewest turns by default)
  │
  ▼
build_sft               full solve history → messages; judge labels, difficulty
                        report and final-eval stats → metadata

Roles

Role settings live in the search_agent.roles config section and mirror the original step_models:

Role Purpose Default temperature
strategist Decide next action from state/history/judge report 0.0
synthesis Expand (atomic-QA merge) and Refine (FUZZ) rewriting 0.7
search_sim LLM-simulated web search / browse (mock mode) 0.7
judge Difficulty report, answer equivalence, quality gate 0.0
solver ReAct rollout with web_search / web_browse 0.7
fast_verify Cheap plan-based verification (optional) 0.0

Each role may set model_id, temperature, and max_tokens; unset fields fall back to the backend default model.

Tools

search_agent.tools.mode selects the tool implementation:

  • mock (default): the search_sim role simulates Google-style search results and page contents. Best for data synthesis.
  • real: Google Custom Search API (google_api_key / google_cx) plus the Jina Reader API (jina_api_key, optional) with an optional SQLite cache (cache_db_path). Best for evaluation and real-data trajectories.

Trajectory Format

Trajectories follow the messages convention used across easydistill agent pipelines: assistant turns contain brief reasoning plus a <tool_call>{"name": ..., "arguments": ...}</tool_call> block, tool outputs return as user turns wrapped in <tool_response>...</tool_response>, and the final assistant turn terminates with <answer>...</answer>.

Usage

easydistill --config configs/pipeline/search_agent_distill_pai_token.yaml

Seed rows accept question/q/instruction and answer/a_star/short_answer aliases. See examples/seed_search_qa.jsonl and configs/pipeline/search_agent_distill_pai_token.yaml (or the _pai_eas variant) for a complete configuration.

Output Sample

Each SFT sample carries the full multi-turn conversation in messages and a self-contained audit trail in metadata: task/seed provenance, hops, the judge difficulty report, per-run correctness, and the final-eval statistics (accuracy, avg_turns). Filtered tasks never reach the SFT stage; set keep_filtered: true on search_task_evolve to keep them in intermediate outputs for debugging.