Dutch YouTube Interface Localization and Content Management

tastewise · Media & Entertainment · 2024 · source Agentic: no · promotional bias: med · source-backed · extracted with prompt 0.5.6 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Food and beverage market intelligence

Tastewise needed non-technical users to ask natural-language food-and-beverage questions and get answers from its proprietary data platform without learning page navigation, filters, API fields, or internal AI models.

Constraints:

The system had to translate free-form natural-language questions into structured Tastewise API requests.
Questions could be follow-ups that depend on previous conversational context.
The platform has multiple pages and filters such as query term, location, audience, and date range.
A single large prompt became hard to debug, improve, and make reliable.
The model could hallucinate unsupported locations because it tried to satisfy the user.

Architecture

TasteGPT is a RAG-like natural-language interface over Tastewise's proprietary data and AI models. A Navigator receives the user's question, detects follow-ups, generates a standalone question when needed, chooses the relevant page, extracts terms and filters, and builds a JSON API request. Tastewise's platform returns relevant data, then a Summarizer combines the data with the original or generated question to create the final answer. The team evolved from one large prompt to parallel specialized prompts and added intermediate layers for location extraction and validation.

Diagram confidence: explicit (topology described by the source).

flowchart LR
  user-question[User question] --> navigator[Navigator]
  navigator --> parallel-prompt-workers[Parallel prompt workers]
  parallel-prompt-workers --> intermediate-validation-layer[Intermediate validation layer]
  navigator --> tastewise-api[Tastewise API]
  tastewise-api --> summarizer[Summarizer]

Drag to pan · scroll or use +/− to zoom · reconstructed from the source's described component topology

Components

Component	Role	Interacts with
`User question`	Natural-language query from a Tastewise user.	`Navigator`
`Navigator`	Determines follow-up state, page selection, query terms, filters, and JSON API request.	`Parallel prompt workers`, `Tastewise API`
`Parallel prompt workers`	Specialized prompts for page navigation, query extraction, location, date, audience, and other filters.	`Navigator`
`Intermediate validation layer`	Separates extraction of locations from final relevance/validity decisions.	`Parallel prompt workers`
`Tastewise API`	Proprietary data and AI model platform returning relevant food-and-beverage data.	`Summarizer`
`Summarizer`	Creates final user-facing answers from returned data and question context.	`Tastewise API`

Data flow

The user asks a natural-language question. TasteGPT translates it to JSON for the Tastewise platform, which retrieves relevant data from proprietary databases and AI models. The Navigator decides whether the request is a follow-up, selects the page, and extracts filters; the Summarizer receives the returned data plus the question and writes the answer.

“taste GPT tast GPT needs to answer to understand the question and to translate it somehow to a Json right a way to interact with the taste wise platform”transcript, stated

“we sent the Json into the platform and now we receive the data back okay so we will take the relevant data we picked it we get it”transcript, stated

Why this architecture

Tastewise needed the LLM to interact with its internal data, not merely answer from general knowledge. Splitting responsibilities into parallel prompts made the system easier to improve, debug, and extend. The intermediate location layer reduced hallucinations by letting the model extract mentioned places first and then decide separately whether they were supported Tastewise locations.

“tgpt it's only a tool to interact with our data with our AI models”transcript, stated

“if you have a problem with one of the prompts you can simply disconnect all the other Focus only on this one”transcript, stated

Models used: ChatGPT · TasteGPT

Scale

Tastewise moved from ChatGPT release awareness in December 2022 to a production first version by March 2023.
“by March we add the first version of tgp in production already it was not you know the greatest but it was something that began”transcript, stated
Tastewise launched its second version by June 2023.
“by June 2023 we had the taste way Summit which was again our goal to launch the second version of taste GPT and we made it”transcript, stated
The platform supports multiple markets such as USA and location filters.
“we support multiple markets okay USA is one of them so it's also important to understand to which in which Market we should focus”transcript, stated

Alternatives considered (roads not taken)

Use one magnificent prompt to infer every API field.

Rejected because: The team found that naive single-prompt approach was clearly wrong and difficult to debug or improve.

“we thought that we will get the question from the user and then we will have this you know magnificent prompt”transcript, stated

“that was the approach like after that after this stage we will group together all the results and then we will create the final API uh request”transcript, stated

Run prompt responsibilities sequentially in a pipeline.

Rejected because: The speaker says parallel prompts reduce waiting, isolate failures, and simplify debugging.

“it's running in parallel okay it means that you you you're not supposed to wait for the for the previous prompt”transcript, stated

Ask the final location prompt to both extract and validate locations in one step.

Rejected because: The model hallucinated unsupported countries while trying to satisfy the user.

“Belgium is one of the countries we support which makes it a valid country but as it's been written in this rule it's clear that Belgium is not among the countries”transcript, stated

Patterns

`tool-use-orchestration` (orchestration)

TasteGPT turns natural language into structured JSON API calls against Tastewise's platform.

“translate it to an API request specifically for The Taste wise platform and it will receive all the relevant data from our databases”transcript, stated

`prompt-chaining` (prompting)

Specialized prompts handle distinct subtasks and their outputs are grouped into a final API request.

“we will group together all the results and then we will create the final API uh request again that's basically the idea”transcript, stated

`fallback-degradation` (reliability)

Parallel prompt execution lets one prompt fail without taking down the whole pipeline.

“if one of the prompt is failing again the fact that it's not working one after the other in a pipeline really helps it to not fail”transcript, stated

`structured-output-contract` (data)

The system's intermediate product is a JSON request with explicit page, query, location, audience, and date fields.

“the final API request should look like this we should navigate to the ingredients and flavors page”transcript, stated

Tradeoffs

Prompt architecture

Chose Parallel specialized prompts. over One all-purpose prompt or a strictly sequential chain..

Rationale: Parallel prompts improve latency, debuggability, extensibility, and failure isolation.

“you can add additional prompts and again it won't affect the other so it's really easier to focus this way”transcript, stated

Location extraction design

Chose Intermediate extraction followed by supported-location validation. over Single final-layer location decision..

Rationale: The intermediate layer lets GPT satisfy extraction without forcing unsupported locations into final filters.

“we decided that we need to help GPT somehow so what we came up with is an additional layer”transcript, stated

Natural-language front end

Chose Let users ask free-form questions. over Requiring users to follow Tastewise's expected query patterns..

Rationale: The goal was higher engagement and an intuitive interface.

“if you'll need to ask taste GPT questions based on how taste wise is expecting you to ask it it won't work”transcript, stated

Gotchas (what bit them)

A single prompt was too broad for reliable API construction.

Root cause: The prompt had to choose page, query term, and multiple filters at once.

Resolution: Tastewise split prompt responsibilities and ran them in parallel.

“we thought that we will get the question from the user and then we will have this you know magnificent prompt”transcript, stated

Location filtering hallucinated unsupported places.

Root cause: GPT tried to satisfy the user by placing any recognized country somewhere in the output.

Resolution: Tastewise added an intermediate layer that extracts country/city/state separately before final relevance filtering.

“CH GPT aims to satisfy the user okay you need to understand this this in in many cases may be the reason why you're getting mistakes”transcript, stated

Prompt changes were hard to isolate in the early design.

Root cause: One large prompt or sequential prompt pipeline creates broad coupling between subtasks.

Resolution: Parallel prompts let the team focus on one prompt and add new prompts without affecting the rest.

“you can simply Focus let's say on prompt seven and you can make some changes and it won't affect any other prompt”transcript, stated

Evaluation

Tastewise evaluated the system through production rollout milestones, debugging case studies, prompt-level failure analysis, and business/user engagement impact.

Metrics: First production version by March 2023 · Second version launched by June 2023 · Parallel prompt execution for page, query, location, audience, date, and other filters

“by March we add the first version of tgp in production already it was not you know the greatest but it was something that began”transcript, stated

“it was a huge change it really affected the business the way we speak about taste wise the way our users are using our product”transcript, stated

Outcomes

Moved a generative-AI interface into production by March 2023.
Launched a second TasteGPT version by June 2023.
Made Tastewise data accessible through natural-language questions.
Improved reliability through parallel prompts and intermediate validation layers.

The talk describes a four-month path from February 2023 hackathon goal-setting to a June 2023 second-version launch, plus support for multiple markets and structured filters.

“by February 2023 we had the Genna aaton a taste wise that we basically made a decision that by the end of the aaton we will have something working”transcript, stated

“it was a real a teamwork like it began with the CEO moved forward to the product like everyone were involved sales marketing design”transcript, stated