dingxingdi Skill: bridge-based multi-hop reasoning
1. Capability Definition & Real Case
- Professional Definition: The ability to answer complex queries by traversing a bridge entity, structural patch, or partial sub-clue that links disjointed pieces of evidence across heterogeneous documents. The agent must systematically forage for information 'scents', execute iterative sub-queries based on incomplete partial matches, and build a unified reasoning chain from distinct semantic steps.
- Dimension Hierarchy: Multi-Source Evidence Composition->Cross-Document Synthesis->bridge-based multi-hop reasoning
Real Case
[Case 1]
- Initial Environment: A bounded news corpus containing one article about the Federal Reserve raising interest rates after booming home prices and another article about future Federal Reserve rate decisions depending on incoming economic data. The shared bridge entity is the Federal Reserve.
- Real Question: Does the article from Fortune suggest that the Federal Reserve’s interest rate hikes are a response to past conditions, while The Sydney Morning Herald article indicates future interest rate decisions will be based on incoming economic data?
- Real Trajectory: 1. Note the bridge entity (Federal Reserve) mentioned across both queried documents. 2. Verify the distinct claim anchored in the Fortune article. 3. Verify the distinct claim anchored in the Herald article. 4. Synthesize the joint logical evaluation crossing both bridged claims.
- Real Answer: Yes
- Why this demonstrates the capability: The answer emerges only after the agent notices that both articles talk about the same bridge entity and compares the two disjointed bridged claims. This evaluates fundamental multi-hop traversal without explicit mathematical logic.
[Case 2]
- Initial Environment: An enterprise multi-source repository encompassing draft requirements documents (PRD), an unstructured Slack space, and an employee database.
- Real Question: Find the employee IDs of the author and all reviewers who provided feedback on the CRM analytics PRD.
- Real Trajectory: 1. Retrieve PRD to extract the URL sharing link. 2. Iteratively search Slack channels utilizing the exact URL as an information patch bridge. 3. Extract reviewer names from chat thread replies. 4. Iteratively lookup extracted names resolving to formal 8-character employee IDs.
- Real Answer: Author: eid_c8ebc4b0; Reviewers: eid_e96d2f38, eid_03a183c9, eid_816aea15.
- Why this demonstrates the capability: This requires highly iterative information patch foraging. The agent bridges a formal text to communications tools utilizing semantic markers (URLs, names) executing sequential sub-queries to navigate incomplete partial clues progressively.
[Case 3]
- Initial Environment: A complex gaming lore guide detailing 'Base Game' metrics and 'Expansion Pack' metrics housed in mutually exclusive chunks completely lacking cross-reference markers.
- Real Question: What is the total number of side missions combined in the original game and its expansion pack?
- Real Trajectory: 1. Decompose the multi-part clue into two separate logical bounds (Base vs Expansion). 2. Forage for the expansion count retrieving a chunk supplying exactly a partial hit. 3. Launch a secondary retrieval pass targeting the base parameter specifically. 4. Intersect the distinct foraged outputs successfully.
- Real Answer: 84 side missions (72 in the base game and 12 in the expansion).
- Why this demonstrates the capability: Demonstrates resolution of partial-clue relevance gaps where user intents cannot be fetched instantaneously. The system manages disjointed evidence segments explicitly recognizing when search data constitutes an incomplete truth before proceeding.
Pipeline Execution Instructions
To synthesize data for this capability, you must strictly follow a 3-phase pipeline. Do not hallucinate steps. Read the corresponding reference file for each phase sequentially:
Phase 1: Environment Exploration Read the exploration guidelines to discover raw knowledge seeds:
references/EXPLORATION.mdPhase 2: Trajectory Selection Once Phase 1 is complete, read the selection criteria to evaluate the trajectory:
references/SELECTION.mdPhase 3: Data Synthesis Once a trajectory passes Phase 2, read the synthesis instructions to generate the final data:
references/SYNTHESIS.md