Strong foundation with well-structured tools and comprehensive parameter schemas. All 7 tools have descriptions (avg 180 chars, within baseline 194). All parameters are typed and described. However, output schemas are not explicitly documented in the source code provided, and error handling guidance is minimal. Tool names follow verb_noun convention well (align_theorems, fetch_chunk, resolve_citation, etc.). The server demonstrates sophisticated domain knowledge (NLI clustering, drift detection, citation formatting) but lacks explicit recovery guidance for failure modes.
Cluster statement-like chunks that state the "same theorem" across different books. Two-stage: embedding similarity seeds candidate clusters, then a symmetric NLI verdict confirms each member — so membership is the INTERSECTION of "looks similar" and "entails in both directions", never similarity alone (precision-first; see ADR-015). Each cluster returns a canonical chunk to cite plus a diff of the qualifiers (hypotheses/assumptions) across books, surfacing where two sources state the result under different conditions.
Explore a topic across the corpus — high-recall breadth + coverage. Unlike retrieve (precision-first; abstains on weak evidence), this surfaces BREADTH: the high-recall candidates (no abstain) plus a KG-strata coverage summary ("concept X has N sections across M books; neighbors rich/thin/absent"). Use it when planning a chapter / curriculum across multiple books before drilling into precise retrieve / answer.
Fetch a single chunk record by its UUID. The lowest-level citation primitive: a direct lookup against the index for the chunk a retrieval result pointed at, returned as a full dict (text + all provenance fields). An unknown UUID returns {"found": False} rather than an error — a clean miss, not a failure. Use resolve_citation instead when you also want the exact source line span and a drift check.
Read an exact [start_line, end_line] excerpt from a source file. The verbatim-text primitive: when an agent already knows the file and line range (e.g. from a prior citation), this returns those lines from the book on disk, so quoted prose comes from the source rather than a paraphrase. The read is containment-guarded — a path resolving outside every configured source directory is rejected.
Output schemas not documented in source. While input schemas are comprehensive, return types and field structures are not explicitly declared. LLMs cannot plan downstream tool calls without knowing what fields to expect.
Error handling lacks recovery guidance. Tools return success/failure but do not guide LLMs on what to do next (e.g., 'UUID not found, try search first' or 'Path outside source directory, check source_path parameter'). Errors should be actionable.
format_citation style parameter lacks enum constraint. Accepts arbitrary strings ('APA', 'Chicago', 'MLA', etc.) but does not declare valid options as an enum. LLMs may hallucinate unsupported styles.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 61 | 2026-07-28+ | v2 |
Find equations by rendered text, raw LaTeX, or SymPy tree. Searches the equation index across three views: 'text' (surrounding prose), 'latex' (exact formula body), or 'tree' (canonical operator-tree, collapses notation variants). Returns equation records with parent chunk UUIDs for citation.
Format a citation string from a chunk record. Takes a chunk UUID and optional citation style, returns a formatted citation string suitable for pasting into prose. Supports multiple citation styles (APA, Chicago, MLA, etc.).
Resolve a chunk UUID to a complete citation record with on-disk span verification. Combines fetch_chunk + fetch_span + drift check: given a chunk UUID from retrieval, this returns the chunk record, maps it to its source file and exact line span, performs a SHA-256 drift check against the on-disk file, and renders a canonical citation string. A drift warning is stamped when the source file has been edited after indexing.
explore_topic session_id parameter lacks clear semantics. Description mentions 'R5.3' and 'de-dup chapter-after-chapter' but does not explain what a session_id is, how to obtain one, or when to use it. Opaque to LLMs.
No pagination or result-limiting guidance for high-recall tools. align_theorems and explore_topic return variable-size result sets but do not document max result count or pagination strategy. Large result sets risk context window exhaustion.