AI-Powered Conversational Search Assistant for B2B Foodservice Operations

tyson_foods · Other · 2025 · source Agentic: yes · promotional bias: med · source-backed · extracted with prompt 0.5.4 (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: B2B foodservice product discovery

Tyson Foodservice needed to serve more than 1 million unattended operators who buy through distributors, while making product search understand culinary intent rather than exact catalog wording.

Constraints:

Foodservice operators use professional terms that differ from catalog labels
Tyson sells primarily through distributors rather than direct end-consumer channels
The assistant must stay grounded in Tyson products, recipes, distributors, and relevant website functions

Architecture

Tyson built two connected systems: semantic search over product, recipe, ingredient, and article content using Titan embeddings and OpenSearch Serverless, and a Claude 3.5 Sonnet plus LangGraph assistant deployed on ECS/Fargate. The assistant uses an agent node and tool execution node to call search, distributor, purchasing, promotion, feedback, and high-value-action tools while RDS stores interaction signals for analytics.

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

flowchart LR
  website-search-interface["website-search-interface"]
  semantic-search-service["semantic-search-service"]
  titan-embedding-model["titan-embedding-model"]
  opensearch-serverless-index["opensearch-serverless-index"]
  content-ingestion-cluster["content-ingestion-cluster"]
  agent-node["agent-node"]
  tool-execution-node["tool-execution-node"]
  agent-tools-layer["agent-tools-layer"]
  high-value-action-store["high-value-action-store"]
  website-search-interface --> semantic-search-service
  website-search-interface --> agent-node
  semantic-search-service --> titan-embedding-model
  semantic-search-service --> opensearch-serverless-index
  titan-embedding-model --> opensearch-serverless-index
  content-ingestion-cluster --> titan-embedding-model
  content-ingestion-cluster --> opensearch-serverless-index
  agent-node --> tool-execution-node
  tool-execution-node --> agent-tools-layer
  tool-execution-node --> agent-node
  agent-tools-layer --> opensearch-serverless-index
  agent-tools-layer --> high-value-action-store

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

Components

Component	Role	Interacts with
`website-search-interface`	Tyson Foodservice website search bar and AI assistant interface for operators.	`semantic-search-service`, `agent-node`
`semantic-search-service`	ECS/Fargate search application exposed through an ALB and WAF.	`titan-embedding-model`, `opensearch-serverless-index`
`titan-embedding-model`	Amazon Titan Text Embeddings model used to create vectors for user queries and indexed content.	`opensearch-serverless-index`
`opensearch-serverless-index`	Vector search store for products, recipes, ingredients, and articles.	`semantic-search-service`, `agent-tools-layer`
`content-ingestion-cluster`	Separate ECS/Fargate ingestion process that converts external JSON content into embeddings and loads OpenSearch.	`titan-embedding-model`, `opensearch-serverless-index`
`agent-node`	Claude 3.5 Sonnet prompt node that decides when conversation requires a tool call.	`tool-execution-node`
`tool-execution-node`	Generic LangGraph tool executor that runs Python-function tools and returns their outputs to the agent.	`agent-tools-layer`, `agent-node`
`agent-tools-layer`	Thin wrapper around services and databases for product search, distributors, purchasing, promotions, and feedback.	`opensearch-serverless-index`, `high-value-action-store`
`high-value-action-store`	RDS-backed analytics store for meaningful customer interactions and business-intelligence events.	—

Data flow

Website search queries are embedded with Amazon Titan Text Embeddings, matched by k-NN search in OpenSearch Serverless, and returned to the site. Conversational queries go to a Claude 3.5 Sonnet agent node; when a request needs data or action, LangGraph routes tool calls through a generic executor to service/database tools, then returns summarized results. Ingestion separately embeds JSON content into OpenSearch, and high-value tool events are persisted for analytics.

3 supporting quotes

“The search application performs a k-nearest neighbors (k-NN) vector search to find relevant results in”article, stated

“The user uses the AI assistant interface to ask questions in natural language.”article, stated

“Whenever a tool call is issued by the agent node, this node handles the execution of the tool call.”article, stated

Why this architecture

Tyson separated semantic retrieval from conversational orchestration because the website needed both exact product discovery and multi-turn assistance without disrupting the distributor sales model. Curated embeddings reduce index noise, OpenSearch Serverless lowers search operations overhead, and LangGraph tools let the assistant expose existing business functions while capturing useful demand signals.

“By curating what content gets indexed rather than including everything verbatim, we dramatically improved search relevance while reducing index bloat”article, stated

“We used LangGraph to streamline the implementation process, because it provides several convenient primitives specifically designed for building agentic systems with LLMs.”article, stated

Models used: Amazon Titan Text Embeddings V2 · Anthropic Claude 3.5 Sonnet

Scale

Tyson Foods produces approximately 20% of U.S. beef, pork, and chicken.
“Tyson Foods produces approximately 20% of the nation’s beef, pork, and chicken”article, stated
The assistant targets more than 1 million previously unattended operators.
“Tyson had limited direct engagement with over 1 million unattended operators who purchased their products through distributors without direct company relationships.”article, stated
OpenSearch Serverless scales compute and storage with query volume and catalog size.
“This solution automatically scales compute and storage resources to match query volume and product catalog size without requiring dedicated administrative overhead.”article, stated

Alternatives considered (roads not taken)

Keyword-only product search

Rejected because: Keyword matching missed foodservice terminology variations such as pulled vs. shredded chicken and wings vs. party wings or drummettes.

“Traditional keyword-based search on CPG websites like Tyson Foodservice often falters when customers search for products using industry terminology that varies from official catalog descriptions.”article, stated

Indexing all page copy verbatim

Rejected because: The ingestion workflow uses LLM preprocessing to keep search-critical content and avoid index bloat.

“This approach made sure that purely presentational website copy and non-essential informational text were filtered out”article, stated

Direct-to-consumer checkout inside the assistant

Rejected because: The assistant preserves the distributor model by helping users identify distributors, check availability, or contact sales representatives.

“Tyson foods mainly sells their products through distributors and does not sell them directly.”article, stated

Post-hoc analytics over chat logs

Rejected because: High-value action capture is integrated directly into tool invocation so insight is gathered during useful interactions.

“Rather than analyzing chat logs after the fact, we integrated the capture mechanism directly into the AI assistant’s operational workflow through LangGraph”article, stated

Patterns

`hybrid-search` (product-discovery)

Tyson replaced brittle keyword lookup with semantic vector search that understands culinary terms, preparation methods, and product applications.

“Semantic search transforms this experience by understanding the conceptual relationships between culinary terms, preparation methods, and product applications.”article, stated

`query-rewriting` (index-preprocessing)

LLMs convert product metadata into concise search strings tailored for vector indexing instead of embedding all website copy verbatim.

“We used large language models (LLMs) to analyze and extract only the most relevant elements from each content piece, creating meaningful search strings specifically designed for semantic indexing.”article, stated

`tool-use-orchestration` (conversational-assistant)

Claude uses LangGraph tool calls to retrieve product, recipe, distributor, purchasing, promotion, and feedback information.

“The system uses the tool calling capabilities of Anthropic’s Claude to implement the AI assistant’s agentic behavior.”article, stated

`structured-output-contract` (semantic-indexing)

The indexing prompt requires concise generated search strings enclosed in XML tags and restricted to foodservice-relevant data.

“create a single, concise search string suitable for indexing in a vector database”article, stated

“Enclose the generated search string within XML tags.”article, stated

`observability-tracing` (business-intelligence)

The assistant captures high-value actions during tool execution so natural conversations become structured customer-interest signals.

“we integrated the capture mechanism directly into the AI assistant’s operational workflow through LangGraph, allowing for real-time insight collection during customer interactions.”article, stated

Tradeoffs

How to learn from unattended operators.

Chose capture high-value actions during useful assistant interactions over only relying on page views, clicks, and time-on-site analytics.

Rationale: Conversational tool calls reveal product interests, pain points, purchase intent, and regional needs in a more structured way than conventional web analytics.

“Whereas conventional analytics tools track user behavior through page views, clicks, and time-on-site metrics, our solution uses the rich conversational data generated through natural dialogue.”article, stated

How to preserve the distributor sales model.

Chose help users identify distributors and sales reps over turning the website assistant into a direct-sales checkout.

Rationale: Tyson's B2B model depends on distributors, so the assistant guides customers through distributor availability and sales-rep contact rather than bypassing the channel.

“Tyson foods mainly sells their products through distributors and does not sell them directly.”article, stated

Gotchas (what bit them)

Keyword search missed foodservice terminology variations.

Root cause: Chefs and operators search with terms like pulled chicken or buffalo-style appetizers that may not match exact catalog descriptions.

Resolution: Semantic search maps culinary concepts and product applications to relevant products even when wording differs.

“Chefs searching for “pulled chicken” might miss relevant products labeled as “shredded chicken,” or those looking for “wings” might not see results for “party wings” or “drummettes.””article, stated

Embedding all page copy would bloat and dilute the search index.

Root cause: Presentational website text and non-essential information can overwhelm search-critical product attributes.

Resolution: LLM preprocessing filters to culinary applications, preparation methods, and ingredient specifications before vector indexing.

“purely presentational website copy and non-essential informational text were filtered out, and search-critical elements like culinary applications, preparation methods”article, stated

Assistant tool outputs should not be over-interpreted.

Root cause: Catalog and distributor queries need concrete information from Tyson systems, not free-form model speculation.

Resolution: The prompt tells the assistant to show search results as-is and stay inside product, recipe, distributor, and purchasing boundaries.

“Do not try to intrepret the results of the search tool and show them as it is”article, stated

Evaluation

Tyson evaluates through search relevance, semantic match quality, high-value action logs, customer-interest signals, and the assistant's ability to connect operators to products, distributors, and sales channels.

Metrics: semantic search relevance · high-value action counts · product/category interest · distributor inquiries · bulk purchase or promotion interest

“The system identifies and logs specific high-value interactions whenever users request detailed product information, inquire about specific product categories”article, stated

“Tyson can use these insights to identify trending product categories and potential gaps in their portfolio”article, stated

Outcomes

Operators can search by culinary intent rather than exact catalog terms.
Tyson gains a direct website interface for previously unattended foodservice operators.
The assistant captures conversational business-intelligence signals as part of normal service.
The architecture exposes search, distributor, purchasing, promotion, and feedback functionality through natural language.

Tyson Foodservice targets more than 1 million unattended operators and draws from product, recipe, ingredient, article, distributor, and customer-interaction data.

“Until recently, Tyson had limited direct engagement with over 1 million unattended operators who purchased their products through distributors without direct company relationships.”article, stated

“By transforming every customer conversation into structured, actionable data, Tyson Foodservice can now measure customer interest with unprecedented precision”article, stated