Solago CI Dashboard

One dashboard for the whole competitive picture.

Solago is Germany's largest online photovoltaics retailer, and they needed to see the whole battlefield in one place: what customers say about them and eight competitors, how their product catalog stacks up, and where prices are moving. The platform pulls reviews from three sources twice a day and turns them into monitoring, sentiment analysis, a knowledge bot and a price tool.

Grounding, hybrid search, and an eval harness.

The product advisor had to be trustworthy in a domain where a wrong price or a made-up product is a real problem — so most of the engineering went into making it faithful:

• Hybrid search fuses dense vectors (Pinecone) with Postgres full-text via reciprocal-rank fusion, with a score threshold and a low-confidence flag instead of a confident guess.
• A strict grounding prompt only lets the model use product names and prices that appear verbatim in tool results — "every extra product is a serious error."
• Answers render as generative UI — product cards, comparison tables — not markdown, and the grounding rules apply to the components too.
• A RAG eval harness with 50 golden queries and an LLM-as-judge measures hit-rate and faithfulness, so quality is a number, not a vibe.

Sentiment, fake-review detection, price tracking.

The platform goes well past search:

• Incremental sentiment analysis extracts topics, pros and cons, and a fake-review probability from every review.
• A price tool detects the shop platform (Shopify, Shopware, JTL, WooCommerce) and matches products across competitors by model number and EAN.
• Alerts run on a schedule with an email digest, and a documented false-positive analysis turned a messy matching problem into a repeatable playbook.
• Full observability via Langfuse, with a user feedback loop wired straight into the traces.