FindFast

Inspiration

E-commerce conversational agents often struggle with multi-turn shopping journeys, failure to retain conversational context, and stumble when users change their minds, or asking redundant clarifying questions. We built FindFast to create an ultra-fast, high-precision conversational retrieval engine that accurately tracks multi-turn slot constraints, dynamically pivots on user intent overrides, and returns ranked product recommendations with zero LLM inference latency.

What it does

Fine Fast powers an intelligent conversational product discovery pipeline designed for low latency and high retrieval relevance (MRR):

Maintains cumulative user constraints, slot entities (gender, size mappings), and profile preferences across multi-turn sessions.
Identifies conversational pivots (e.g., "actually", "changed my mind", "instead") to flush stale constraints and adapt search focus immediately.
Skips redundant questions by checking accumulated session terms against attribute keyword taxonomies before asking the user for missing details.
Uses exact consecutive phrase (bigram) matching and custom field-weighted scoring across product titles, categories, features, and descriptions.

How we built it

We engineered FindFast entirely in Python as a high-speed conversational retrieval engine designed for zero-inference latency. We implemented a lightweight state machine in Python to maintain cumulative session constraints, handle entity slot mappings, and dynamically flush stale parameters upon detecting intent pivots (such as "actually" or "instead"). The retrieval core avoids slow LLM API calls by using custom Python-based tokenization, consecutive bigram matching, taxonomy lookups to prune redundant questions, and field-weighted ranking heuristics across product metadata to return high-MRR results in milliseconds.

Challenges we ran into

Our biggest hurdle was navigating heuristic iteration and score regressions, where several intuitively promising algorithmic ideas actually lowered our retrieval benchmark scores rather than improving them. This forced us to continuously test, reject, and rethink new scoring strategies through rapid data-driven trial and error, while simultaneously engineering strict state boundary handling to accurately differentiate between additive constraints and total intent overrides without polluting search vectors.

Accomplishments that we're proud of

We successfully built a robust state machine that reliably adapts to intent pivots, coupled with a custom field-weighted scoring algorithm that consistently delivers top-tier relevance and surfaces the most accurate candidate products directly to the user.

What we learned

Building FindFast taught us that engineering intuition must always yield to rigorous empirical validation, as seemingly obvious ranking heuristics frequently degrade benchmark metrics like MRR in real search pipelines. We also gained a deep appreciation for the power of deterministic NLP, realizing that well-crafted taxonomies, token state tracking, and n-gram scoring can deliver vastly superior speed, precision, and cost efficiency over heavy generative models for structured multi-turn search.

What's next for FindFast

Next, we want to bring FindFast directly into TikTok Shop as an ultra-fast, in-app shopping assistant that keeps up with fast-paced social commerce. We plan to expand our engine to recognize trending slang and viral aesthetics (like "clean girl look" or "gym pump covers"), automatically organize messy merchant catalogs and product tags so items don't get lost, and dynamically rerank search results based on the videos, livestreams, and products users engage with in real time.