Cheap retrieval-only field orientation (TOOL-06, v3 get_field_orientation tool backend).

Embeds the topic via Gemini Flash (get_query_embedding), runs a cosine search over the papers vector index to fetch candidate papers, then re-ranks those candidates by a foundational-ness blend (0.6 * normalised citation_count + 0.4 * cosine_similarity). Returns the top limit papers as a thin list with similarity score and citation count.

THIS IS NOT THE FIELD GUIDE AGENT. No DeepSeek synthesis, no quota, no Pro requirement, no background jobs. The expensive LLM synthesis half stays in the /field-guide skill (docs/architecture/tool-vs-skill.md §field_guide row) which composes this retrieval endpoint + in-context reasoning client-side.

Rate limit: 20/minute per API key.

Response shape: topic: str papers: list[{arxiv_id, title, authors, published_date, citation_count, rank_score, llm_summary, llm_novelty_score, similarity}] total: int note: str — reminder that this is retrieval-only, no LLM synthesis