Full-cycle AI fashion assistant with agent brain over a clean fashion knowledge core. Provides styling, brand DNA analysis, trend forecasting, and personal stylist capabilities via FastAPI.
FashionGraph exposes 18 tools with significant definition gaps. Most tools lack proper input schemas (parameters missing type definitions or descriptions). Tool names are inconsistent: 'search' appears twice with different signatures (web search vs. embedding search), violating the single-responsibility principle. Descriptions exist but are often vague (e.g., 'Personal styling advice' without explaining when to use it vs. 'brand' tool). Critical tools like 'agent' and 'chat' accept complex nested objects (selection array, learn boolean) but lack parameter-level descriptions. No output schemas are documented. Error handling is absent, no guidance on retryability, user-fixable errors, or recovery paths. Security concerns: 'add_document' and 'add_expert_annotation' accept arbitrary metadata dicts without validation. Vision tools ('save', 'load', 'build_product_index') expose file paths as parameters, risking path traversal. Overall, the server reads as a research prototype rather than production-grade tooling.
Splits text into chunks and upserts them into ChromaDB for RAG retrieval
Adds an expert trend annotation to the knowledge base
The autonomous ReAct agent router: styling, brand DNA, trends, lineage, live facts. Returns answer + canvas cards + sources
Reviews an uploaded outfit photo (garments, styling, lineage, sources)
Define a brand's DNA (aesthetic, values, materials, palette)
Builds a visual index from the product-image dataset by streaming product rows, decoding images, and embedding them in batches
Alias of /agent for free-form chat
Duplicate tool name 'search' with incompatible signatures: web search (string query) vs. embedding search (array query). LLMs cannot disambiguate. Violates single-responsibility principle.
Parameter 'metadata' in add_document and add_expert_annotation accepts arbitrary dicts with no schema validation. Allows injection of unexpected fields. No type constraints or required-field declarations.
File path parameters ('path' in save/load/build_product_index) exposed directly without validation. Risk of path traversal attacks. No sanitization or directory boundary checks visible.
No output schemas documented for any tool. LLMs cannot plan downstream calls or extract required fields. E.g., 'agent' tool returns 'answer + canvas cards + sources' but structure is undefined.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 43 | <=2025-11-25 | v2 |
Reviews several garments/looks together as one outfit (one vision call)
Liveness check and component status
Facts + designer lineage from the knowledge graph
Loads a visual index from a .npz file
Returns the onboarding questions for brand bootstrapping
Builds the brand foundation from onboarding answers
Saves the visual index to a .npz file
Returns the top-k most similar items to a query embedding
Live web search — current facts, prices, latest shows
Personal styling advice (outfits, occasion, capsule)
Analyse a trend, or rate how plausible/emerging one is
Complex nested parameters ('selection' array in agent/chat, 'request' object in run) lack per-field descriptions. LLMs cannot infer structure or valid values.