Fresh Market Intelligence is a Streamlit app and LangGraph pipeline that transforms official trade data, EU price series, and curated news into a concise, decision-ready market brief for fresh produce.
It focuses on procurement questions:
What is the latest official import baseline (volume, value, price)?
Who are the top origins and how are they changing month to month?
Are there fresh signals (news, logistics, weather, harvest) that affect supply?
What do EU supply chain prices indicate for the selected market?
The core problem - Fresh-produce purchasing managers face information overload with incomplete data:
News coverage is inconsistent by product, geography, and week
Official prices and volumes are delayed, aggregated, or expensive to access
Critical signals are fragmented across trade data, weather effects, and market news
Silence itself can be meaningful, but most systems ignore it
Most AI systems implicitly assume - "If data is missing, we failed." That assumption is wrong for agricultural markets.
The goal is simple - Help purchasing managers make better decisions when market data is incomplete, delayed, or noisy.
COMEXT import summary with top origins, MoM and YoY deltas, and seasonality baseline for the latest official month.
EU Agrifood price explorer (weekly or monthly) with chart, variety breakdown, and a compact price summary.
News collection via Google PSE (optional) + cached pool fallback, recency filtering, relevance filtering, optional full text enrichment (trafilatura), and HITL approval.
Digest generator built with LangGraph that merges baseline + prices + signals and stores the result in SQLite for later retrieval.
Optional LangSmith tracing for step level debugging.
Plans evidence collection for a given market topic
Collects authoritative trade and price baselines
Adds market signals only when they add value
Reasons carefully over missing information
Produces a structured procurement decision brief
Stores results to build long-term market memory
Output - Each run produces a procurement decision brief that includes: key findings, trade baseline, prices (EU API), market signals (if available), explicit data gaps, recommended actions, evidence grade + coverage score, key takeaways, evidence links, notes.
The output is suitable for buyers, category managers, and decision reviews.
The system became reliable only after we made the pipeline explicit and conservative - LangGraph exposes reasoning steps, COMEXT delays/missing data are handled via cached facts + validation so runs don’t break, prompts force uncertainty instead of hallucinations, news is included only when it improves the brief (HITL helps), and safe URL fetching + deterministic fallbacks keep summaries stable.
Next steps include expanding market coverage, improving signal quality, and refining how uncertainty is communicated to support faster decision reviews.
