EntropyAgent

Inspiration

Modern e-commerce search relies heavily on static keyword matching, which often fails to capture complex user intent or adapt to mid-conversation changes. We built EntropyAgent, a multi-turn shopping agent that turns vague customer requests into precise, top-ranked recommendations within 10 turns—all while running completely offline, without sending customer data to external cloud services.

What it does

Our agent processes multi-turn shopping interactions through a 5-step pipeline:

1. Structured Memory & Intent Override Handling

Tracks hard constraints, declined preferences, and previously shown products. If a shopper changes their mind, stale constraints are retired automatically.

2. Field-Weighted Lexical Retrieval

Queries a 50,000-product catalog using SQLite FTS5 BM25 with dynamic field weights across titles, categories, features, and details.

3. Catalog-Grounded Concept Reranking

Converts items into up to 16 canonical concepts (e.g., category, material, color, use_case) and reranks candidates using local vector embeddings (when LLM is enabled):

$$\text{Similarity}(q, d) = \frac{\mathbf{q} \cdot \mathbf{d}}{\Vert{}\mathbf{q}\Vert{} \Vert{}\mathbf{d}\Vert{}}$$

4. Information-Gain Clarification

Calculates rank-weighted entropy over the current candidate pool to dynamically select the single most informative attribute question:

$$\text{GainRatio}(A) = \frac{\text{Entropy}(S) - \text{Entropy}_A(S)}{\text{SplitInfo}(A)}$$

5. Grounded Top-10 Output Returns up to 10 valid, unseen product recommendations alongside the next clarification query.

How we built it

• Development Tools: VS Code, Python 3.10+, SQLite FTS5, Codex, Claude Code

• AI & Local Modeling: Ollama (llama3.2:3b for catalog-constrained concept selection and nomic-embed-text-v2-moe for local semantic embeddings)

• Core Algorithms: Custom rank-weighted entropy question selection, field-weighted BM25 lexical indexing, and vector cosine similarity

• Datasets & Assets Used: Amazon Reviews 2023 (50,000-product frozen dataset)

Challenges we ran into

Embedding raw product rows was noisy and slow. We optimized this by generating structured canonical attribute indices (catalog_attributes.jsonl) and embedding unique canonical concepts into SQLite once.

Accomplishments that we're proud of

• Jumped from 12.5% Hit Rate@10 on the baseline to 98.5%, while reducing average resolution speed from 9.8 turns to 3.4 turns.

• Achieved high performance running locally at $0 cloud API cost.

• Designed a deterministic fallback path that operates smoothly even when Ollama is disabled.

What we learned

We gained deep insights into constructing fast SQLite full-text indices, implementing information-gain algorithms for decision trees in conversational recommendation systems, and balancing local embedding quality with real-time response constraints.

We also learned that sometimes, simple is better. Overcomplicating things sometimes lead to worse products.