Databricks Databricks-Generative-AI-Engineer-Associate Databricks Certified Generative AI Engineer Associate Exam Practice Test

The task requires an LLM pipeline for multi-stage reasoning with external tools, necessitating planning, adaptability, and complex reasoning. Let’s evaluate the options based on Databricks’ recommendations for advanced LLM workflows.

Option A: Train the LLM to generate a single, comprehensive response without interacting with any external tools, relying solely on its pre-trained knowledge

This approach limits the LLM to its static knowledge base, excluding external tools and multi-stage reasoning. It can’t adapt or plan actions dynamically, failing the requirements.

Databricks Reference:"External tools enhance LLM capabilities beyond pre-trained knowledge"("Building LLM Applications with Databricks," 2023).

Option B: Implement a framework like ReAct which allows the LLM to generate reasoning traces and perform task-specific actions that leverage external tools if necessary

ReAct (Reasoning + Acting) combines reasoning traces (step-by-step logic) with actions (e.g., tool calls), enabling the LLM to plan, adapt, and execute complex tasks iteratively. This meets all requirements: multi-stage reasoning, tool use, and adaptability.

Databricks Reference:"Frameworks like ReAct enable LLMs to interleave reasoning and external tool interactions for complex problem-solving"("Generative AI Cookbook," 2023).

Option C: Encourage the LLM to make multiple API calls in sequence without planning or structuring the calls, allowing the LLM to decide when and how to use external tools spontaneously

Unstructured, spontaneous API calls lack planning and may lead to inefficient or incorrect tool usage. This doesn’t ensure effective multi-stage reasoning or adaptability.

Databricks Reference: Structured frameworks are preferred:"Ad-hoc tool calls can reduce reliability in complex tasks"("Building LLM-Powered Applications").

Option D: Use a Chain-of-Thought (CoT) prompting technique to guide the LLM through a series of reasoning steps, then manually input the results from external tools for the final answer

CoT improves reasoning but relies on manual tool interaction, breaking automation and adaptability. It’s not a scalable pipeline solution.

Databricks Reference:"Manual intervention is impractical for production LLM pipelines"("Databricks Generative AI Engineer Guide").

Conclusion: Option B (ReAct) is the best approach, as it integrates reasoning and tool use in a structured, adaptive framework, aligning with Databricks’ guidance for complex LLM workflows.

Problem Context: The problem involves matching team members to new projects based on two main factors:

Availability: Ensure the team members are available during the project dates.

Profile-Project Match: Use the employee profiles (unstructured text) to find the best match for a project’s scope (also unstructured text).

The two main inputs are theemployee profilesandproject scopes, both of which are unstructured. This means traditional rule-based systems (e.g., simple keyword matching) would be inefficient, especially when working with large datasets.

Explanation of Options: Let's break down the provided options to understand why D is the most optimal answer.

Option Asuggests embedding project scopes into a vector store and then performing retrieval using team member profiles. While embedding project scopes into a vector store is a valid technique, it skips an important detail: the focus should primarily be on embedding employee profiles because we're matching the profiles to a new project, not the other way around.

Option Binvolves using a large language model (LLM) to extract keywords from the project scope and perform keyword matching on employee profiles. While LLMs can help with keyword extraction, this approach is too simplistic and doesn’t leverage advanced retrieval techniques like vector embeddings, which can handle the nuanced and rich semantics of unstructured data. This approach may miss out on subtle but important similarities.

Option Csuggests calculating a similarity score between each team member's profile and project scope. While this is a good idea, it doesn’t specify how to handle the unstructured nature of data efficiently. Iterating through each member’s profile individually could be computationally expensive in large teams. It also lacks the mention of using a vector store or an efficient retrieval mechanism.

Option Dis the correct approach. Here’s why:

Embedding team profiles into a vector store: Using a vector store allows for efficient similarity searches on unstructured data. Embedding the team member profiles into vectors captures their semantics in a way that is far more flexible than keyword-based matching.

Using project scope for retrieval: Instead of matching keywords, this approach suggests using vector embeddings and similarity search algorithms (e.g., cosine similarity) to find the team members whose profiles most closely align with the project scope.

Filtering based on availability: Once the best-matched candidates are retrieved based on profile similarity, filtering them by availability ensures that the system provides a practically useful result.

This method efficiently handles large-scale datasets by leveragingvector embeddingsandsimilarity searchtechniques, both of which are fundamental tools inGenerative AI engineeringfor handling unstructured text.

Technical References:

Vector embeddings: In this approach, the unstructured text (employee profiles and project scopes) is converted into high-dimensional vectors using pretrained models (e.g., BERT, Sentence-BERT, or custom embeddings). These embeddings capture the semantic meaning of the text, making it easier to perform similarity-based retrieval.

Vector stores: Solutions likeFAISSorMilvusallow storing and retrieving large numbers of vector embeddings quickly. This is critical when working with large teams where querying through individual profiles sequentially would be inefficient.

LLM Integration: Large language models can assist in generating embeddings for both employee profiles and project scopes. They can also assist in fine-tuning similarity measures, ensuring that the retrieval system captures the nuances of the text data.

Filtering: After retrieving the most similar profiles based on the project scope, filtering based on availability ensures that only team members who are free for the project are considered.

This system is scalable, efficient, and makes use of the latest techniques inGenerative AI, such as vector embeddings and semantic search.

Problem Context: When using data to train a model, it’s essential to ensure compliance with licensing to avoid legal risks. Legal issues can arise from using data without permission, especially when it comes from third-party sources.

Explanation of Options:

Option A: Reaching out to data curatorsbeforeusing the data is an appropriate action. This allows you to ensure you have permission or understand the licensing terms before starting to use the data in your model.

Option B: Usingoriginal datathat you personally created is always a safe option. Since you have full ownership over the data, there are no legal risks, as you control the licensing.

Option C: Using data that is explicitly labeled with an open license and adhering to the license terms is a correct and recommended approach. This ensures compliance with legal requirements.

Option D: Reaching out to the data curatorsafteryou have already started using the trained model isnot appropriate. If you’ve already used the data without understanding its licensing terms, you may have already violated the terms of use, which could lead to legal complications. It’s essential to clarify the licensing termsbeforeusing the data, not after.

Thus,Option Dis not appropriate because it could expose you to legal risks by using the data without first obtaining the proper licensing permissions.

Addressing offensive or inflammatory outputs in a Retrieval-Augmented Generation (RAG) system is critical for improving user experience and ensuring ethical AI deployment. Here's whyDis the most effective approach:

Manual data curation: The root cause of offensive outputs often comes from the underlying data used to train the model or populate the retrieval system. By manually curating the upstream data and conducting thorough reviews before the data is fed into the RAG system, the engineer can filter out harmful, offensive, or inappropriate content.

Improving data quality: Curating data ensures the system retrieves and generates responses from a high-quality, well-vetted dataset. This directly impacts the relevance and appropriateness of the outputs from the RAG system, preventing inflammatory content from being included in responses.

Effectiveness: This strategy directly tackles the problem at its source (the data) rather than just mitigating the consequences (such as informing users or restricting access). It ensures that the system consistently provides non-offensive, relevant information.

Other options, such as increasing the frequency of data updates or informing users about behavior expectations, may not directly mitigate the generation of inflammatory outputs.

The task is to deploy an LLM application using Foundation Model APIs in a production environment while adhering to security best practices. Authentication is critical for securing access to Databricks resources, such as the Foundation Model API. Let’s evaluate the options based on Databricks’ security guidelines for production scenarios.

Option A: Use an access token belonging to service principals

Service principals are non-human identities designed for automated workflows and applications in Databricks. Using an access token tied to a service principal ensures that the authentication is scoped to the application, follows least-privilege principles (via role-based access control), and avoids reliance on individual user credentials. This is a security best practice for production deployments.

Databricks Reference:"For production applications, use service principals with access tokens to authenticate securely, avoiding user-specific credentials"("Databricks Security Best Practices," 2023). Additionally, the "Foundation Model API Documentation" states:"Service principal tokens are recommended for programmatic access to Foundation Model APIs."

Option B: Use a frequently rotated access token belonging to either a workspace user or a service principal

Frequent rotation enhances security by limiting token exposure, but tying the token to a workspace user introduces risks (e.g., user account changes, broader permissions). Including both user and service principal options dilutes the focus on application-specific security, making this less ideal than a service-principal-only approach. It also adds operational overhead without clear benefits over Option A.

Databricks Reference:"While token rotation is a good practice, service principals are preferred over user accounts for application authentication"("Managing Tokens in Databricks," 2023).

Option C: Use OAuth machine-to-machine authentication

OAuth M2M (e.g., client credentials flow) is a secure method for application-to-service communication, often using service principals under the hood. However, Databricks’ Foundation Model API primarily supports personal access tokens (PATs) or service principal tokens over full OAuth flows for simplicity in production setups. OAuth M2M adds complexity (e.g., managing refresh tokens) without a clear advantage in this context.

Databricks Reference:"OAuth is supported in Databricks, but service principal tokens are simpler and sufficient for most API-based workloads"("Databricks Authentication Guide," 2023).

Option D: Use an access token belonging to any workspace user

Using a user’s access token ties the application to an individual’s identity, violating security best practices. It risks exposure if the user leaves, changes roles, or has overly broad permissions, and it’s not scalable or auditable for production.

Databricks Reference:"Avoid using personal user tokens for production applications due to security and governance concerns"("Databricks Security Best Practices," 2023).

Conclusion: Option A is the best choice, as it uses a service principal’s access token, aligning with Databricks’ security best practices for production LLM applications. It ensures secure, application-specific authentication with minimal complexity, as explicitly recommended for Foundation Model API deployments.

Problem Context: The Generative AI Engineer needs a prompt that will enable an LLM trained on customer call transcripts to classify and respond correctly regarding product availability. The desired response should clearly indicate whether a product is "In Stock" or "Out of Stock," and it should be formatted in a way that is structured and easy to parse programmatically, such as JSON.

Explanation of Options:

Option A: Respond with “In Stock” if the customer asks for a product. This prompt is too generic and does not specify how to handle the case when a product is not available, nor does it provide a structured output format.

Option B: This option is correctly formatted and explicit. It instructs the LLM to respond based on the availability mentioned in the customer call transcript and to format the response in JSON. This structure allows for easy integration into systems that may need to process this information automatically, such as customer service dashboards or databases.

Option C: Respond with “Out of Stock” if the customer asks for a product. Like option A, this prompt is also insufficient as it only covers the scenario where a product is unavailable and does not provide a structured output.

Option D: While this prompt correctly specifies how to respond based on product availability, it lacks the structured output format, making it less suitable for systems that require formatted data for further processing.

Given the requirements for clear, programmatically usable outputs,Option Bis the optimal choice because it provides precise instructions on how to respond and includes a JSON format example for structuring the output, which is ideal for automated systems or further data handling.

 Problem Context: The Generative AI Engineer needs a model for a Retrieval-Augmented Generation (RAG) application that provides high-quality answers, where latency and throughput are not major concerns. The key factors areconfidentialityandsensitivityof the data, as well as the requirement for all processing to be confined to internal resources without external data transmission.

 Explanation of Options:

Option A: Dolly 1.5B: This model does not typically support RAG applications as it's more focused on image generation tasks.

Option B: OpenAI GPT-4: While GPT-4 is powerful for generating responses, its standard deployment involves cloud-based processing, which could violate the confidentiality requirements due to external data transmission.

Option C: BGE-large: The BGE (Big Green Engine) large model is a suitable choice if it is configured to operate on-premises or within a secure internal environment that meets regulatory requirements. Assuming this setup, BGE-large can provide high-quality answers while ensuring that data is not transmitted to third parties, thus aligning with the project's sensitivity and confidentiality needs.

Option D: Llama2-70B: Similar to GPT-4, unless specifically set up for on-premises use, it generally relies on cloud-based services, which might risk confidential data exposure.

Given the sensitivity and confidentiality concerns,BGE-largeis assumed to be configurable for secure internal use, making it the optimal choice for this scenario.

The task involves building a question-answering chatbot for an automotive company using a database of vehicle-related documents. The chatbot must efficiently process customer inquiries and provide accurate responses. Let’s evaluate each component to determine which isnotuseful, per Databricks Generative AI Engineer principles.

Option A: Response-generating LLM

An LLM is essential for generating natural language responses to customer queries based on retrieved information. This is a core component of any chatbot.

Databricks Reference:"The response-generating LLM processes retrieved context to produce coherent answers"("Building LLM Applications with Databricks," 2023).

Option B: Invite users to submit long, rather than concise, questions

Encouraging long questions is a user interaction design choice, not a technical component of the chatbot’s architecture. Moreover, long, verbose questions can complicate intent detection and retrieval, reducing efficiency and accuracy—counter to best practices for chatbot design. Concise questions are typically preferred for clarity and performance.

Databricks Reference: While not explicitly stated, Databricks’ "Generative AI Cookbook" emphasizes efficient query processing, implying that simpler, focused inputs improve LLM performance. Inviting long questions doesn’t align with this.

Option C: Vector database

A vector database stores embeddings of the vehicle documents, enabling fast retrieval of relevant information via semantic search. This is critical for a question-answering system with a large document corpus.

Databricks Reference:"Vector databases enable scalable retrieval of context from large datasets"("Databricks Generative AI Engineer Guide").

Option D: Embedding model

An embedding model converts text (documents and queries) into vector representations for similarity search. It’s a foundational component for retrieval-augmented generation (RAG) in chatbots.

Databricks Reference:"Embedding models transform text into vectors, facilitating efficient matching of queries to documents"("Building LLM-Powered Applications").

Conclusion: Option B is not a usefulcomponentin building the chatbot. It’s a user-facing suggestion rather than a technical building block, and it could even degrade performance by introducing unnecessary complexity. Options A, C, and D are all integral to a Databricks-aligned chatbot architecture.

In this scenario, the chatbot must avoid answering political questions and instead provide a standard message for such inquiries. Implementing aSafety Guardrailis the appropriate solution for this:

What is a Safety Guardrail?Safety guardrails are mechanisms implemented in Generative AI systems to ensure the model behaves within specific bounds. In this case, it ensures the chatbot does not answer politically sensitive or irrelevant questions, which aligns with the business rules.

Preventing Responses to Political Questions:The Safety Guardrail is programmed to detect specific types of inquiries (like political questions) and prevent the model from generating responses outside its intended domain. When such queries are detected, the guardrail intervenes and provides a pre-defined response: “Sorry, I cannot answer that. I am a chatbot that can only answer questions around insurance.”

How It Works in Practice:The LLM system can include aclassification layeror trigger rules based on specific keywords related to politics. When such terms are detected, the Safety Guardrail blocks the normal generation flow and responds with the fixed message.

Why Other Options Are Less Suitable:

B (Security Guardrail): This is more focused on protecting the system from security vulnerabilities or data breaches, not controlling the conversational focus.

C (Contextual Guardrail): While context guardrails can limit responses based on context, safety guardrails are specifically about ensuring the chatbot stays within a safe conversational scope.

D (Compliance Guardrail): Compliance guardrails are often related to legal and regulatory adherence, which is not directly relevant here.

Therefore, aSafety Guardrailis the right framework to ensure the chatbot only answers insurance-related queries and avoids political discussions.

Problem Context: The task is to design responses for a healthcare-focused chatbot that appropriately addresses the urgency of a patient's symptoms.

Explanation of Options:

Option A: Suggesting articles might be suitable for less urgent inquiries but is inappropriate for symptoms that could indicate a serious condition.

Option B: Given the description of severe symptoms like headaches and dizziness, directing the patient to emergency services is prudent. This aligns with medical guidelines that recommend immediate professional attention for such severe symptoms.

Option C: Offering well-wishes does not address the potential seriousness of the symptoms and lacks appropriate action.

Option D: While gathering more information is part of a detailed assessment, the immediate need here suggests a more urgent response.

Given the potential severity of the described symptoms,Option Bis the most appropriate, ensuring the chatbot directs patients to seek urgent care when needed, potentially saving lives.

When deploying an LLM application for customer service inquiries, the primary focus is on measuring the operational efficiency and quality of the responses. Here's whyAis the correct metric:

Number of customer inquiries processed per unit of time: This metric tracks the throughput of the customer service system, reflecting how many customer inquiries the LLM application can handle in a given time period (e.g., per minute or hour). High throughput is crucial in customer service applications where quick response times are essential to user satisfaction and business efficiency.

Real-time performance monitoring: Monitoring the number of queries processed is an important part of ensuring that the model is performing well under load, especially during peak traffic times. It also helps ensure the system scales properly to meet demand.

Why other options are not ideal:

B. Energy usage per query: While energy efficiency is a consideration, it is not the primary concern for a customer-facing application where user experience (i.e., fast and accurate responses) is critical.

C. Final perplexity scores for the training of the model: Perplexity is a metric for model training, but it doesn't reflect the real-time operational performance of an LLM in production.

D. HuggingFace Leaderboard values for the base LLM: The HuggingFace Leaderboard is more relevant during model selection and benchmarking. However, it is not a direct measure of the model's performance in a specific customer service application in production.

Focusing on throughput (inquiries processed per unit time) ensures that the LLM application is meeting business needs for fast and efficient customer service responses.

To optimize a chunking strategy for a Retrieval-Augmented Generation (RAG) application, the Generative AI Engineer needs a structured approach to evaluating the chunking strategy, ensuring that the chosen configuration retrieves the most relevant information and leads to accurate and coherent LLM responses. Here's whyCandEare the correct strategies:

Strategy C: Evaluation Metrics (Recall, NDCG)

Define an evaluation metric: Common evaluation metrics such as recall, precision, or NDCG (Normalized Discounted Cumulative Gain) measure how well the retrieved chunks match the user's query and the expected response.

Recallmeasures the proportion of relevant information retrieved.

NDCGis often used when you want to account for both the relevance of retrieved chunks and the ranking or order in which they are retrieved.

Experiment with chunking strategies: Adjusting chunking strategies based on text structure (e.g., splitting by paragraph, chapter, or a fixed number of tokens) allows the engineer to experiment with various ways of slicing the text. Some chunks may better align with the user's query than others.

Evaluate performance: By using recall or NDCG, the engineer can methodically test various chunking strategies to identify which one yields the highest performance. This ensures that the chunking method provides the most relevant information when embedding and retrieving data from the vector store.

Strategy E: LLM-as-a-Judge Metric

Use the LLM as an evaluator: After retrieving chunks, the LLM can be used to evaluate the quality of answers based on the chunks provided. This could be framed as a "judge" function, where the LLM compares how well a given chunk answers previous user queries.

Optimize based on the LLM's judgment: By having the LLM assess previous answers and rate their relevance and accuracy, the engineer can collect feedback on how well different chunking configurations perform in real-world scenarios.

This metric could be a qualitative judgment on how closely the retrieved information matches the user's intent.

Tune chunking parameters: Based on the LLM's judgment, the engineer can adjust the chunk size or structure to better align with the LLM's responses, optimizing retrieval for future queries.

By combining these two approaches, the engineer ensures that the chunking strategy is systematically evaluated using both quantitative (recall/NDCG) and qualitative (LLM judgment) methods. This balanced optimization process results in improved retrieval relevance and, consequently, better response generation by the LLM.

Problem Context: The chatbot must handle various types of queries and intelligently route them to the appropriate responses or systems.

Explanation of Options:

Option A: Limiting the chatbot to only previous event information restricts its utility and does not meet the broader business requirements.

Option B: Having two separate chatbots could unnecessarily complicate user interaction and increase maintenance overhead.

Option C: Implementing a multi-step workflow where the chatbot first identifies the type of question and then routes it accordingly is the most efficient and scalable solution. This approach allows the chatbot to handle a variety of queries dynamically, improving user experience and operational efficiency.

Option D: Focusing solely on payments would not satisfy all the specified user interaction needs, such as inquiring about event details.

Option Coffers a comprehensive workflow that maximizes the chatbot’s utility and responsiveness to different user needs, aligning perfectly with the business requirements.

The task at hand is to improve the LLM’s ability to generate poem-like article summaries with the desired tone and style. Using aneutralizerto normalize the tone and style of the underlying documents (option B) will not help improve the LLM’s ability to generate the desired poetic style. Here’s why:

Neutralizing Underlying Documents:A neutralizer aims to reduce or standardize the tone of input data. However, this contradicts the goal, which is to generate text with aspecific tone and style(like haikus). Neutralizing the source documents will strip away the richness of the content, making it harder for the LLM to generate creative, stylistic outputs like poems.

Why Other Options Improve Results:

A (Explicit Instructions in the Prompt): Directly instructing the LLM to generate text in a specific tone and style helps align the output with the desired format (e.g., haikus). This is a common and effective technique in prompt engineering.

C (Few-shot Examples): Providing examples of the desired output format helps the LLM understand the expected tone and structure, making it easier to generate similar outputs.

D (Fine-tuning the LLM): Fine-tuning the model on a dataset that contains examples of the desired tone and style is a powerful way to improve the model’s ability to generate outputs that match the target format.

Therefore, using a neutralizer (option B) isnotan effective method for achieving the goal of generating stylized poetic summaries.

The task is to benchmark LLMs for text translation using an evaluation set with known high-quality examples, requiring a performant metric. Let’s evaluate the options.

Option A: ROUGE metric

ROUGE (Recall-Oriented Understudy for Gisting Evaluation) measures overlap between generated and reference texts, primarily for summarization. It’s less suited for translation, where precision and word order matter more.

Databricks Reference:"ROUGE is commonly used for summarization, not translation evaluation"("Generative AI Cookbook," 2023).

Option B: BLEU metric

BLEU (Bilingual Evaluation Understudy) evaluates translation quality by comparing n-gram overlap with reference translations, accounting for precision and brevity. It’s widely used, performant, and appropriate for this task.

Databricks Reference:"BLEU is a standard metric for evaluating machine translation, balancing accuracy and efficiency"("Building LLM Applications with Databricks").

Option C: NDCG metric

NDCG (Normalized Discounted Cumulative Gain) assesses ranking quality, not text generation. It’s irrelevant for translation evaluation.

Databricks Reference:"NDCG is suited for ranking tasks, not generative output scoring"("Databricks Generative AI Engineer Guide").

Option D: RECALL metric

Recall measures retrieved relevant items but doesn’t evaluate translation quality (e.g., fluency, correctness). It’s incomplete for this use case.

Databricks Reference: No specific extract, but recall alone lacks the granularity of BLEU for text generation tasks.

Conclusion: Option B (BLEU) is the best metric for translation evaluation, offering a performant and standard approach, as endorsed by Databricks’ guidance on generative tasks.

Problem Context: The engineer needs an open-source LLM with a large context window to develop an application.

Explanation of Options:

Option A: DistilBERT: While an efficient and smaller version of BERT, DistilBERT does not provide a particularly large context window.

Option B: MPT-30B: This model, while large, is not specified as being particularly notable for its context window capabilities.

Option C: Llama2-70B: Known for its large model size and extensive capabilities, including a large context window. It is also available as an open-source model, making it ideal for applications requiring extensive contextual understanding.

Option D: DBRX: This is not a recognized standard model in the context of large language models with extensive context windows.

Thus,Option C(Llama2-70B) is the best fit as it meets the criteria of having a large context window and being available for open-source use, suitable for developing robust language understanding applications.

 Problem Context: The task is to build an LLM-based question-answering application that integrates new documents frequently with minimal costs and development efforts.

 Explanation of Options:

Option A: Utilizes a prompt and a retriever, with the retriever output being fed into the LLM. This setup is efficient because it dynamically updates the data pool via the retriever, allowing the LLM to provide up-to-date answers based on the latest documents without needing to frequently retrain the model. This method offers a balance of cost-effectiveness and functionality.

Option B: Requires frequent retraining of the LLM, which is costly and labor-intensive.

Option C: Only involves prompt engineering and an LLM, which may not adequately handle the requirement for incorporating new documents unless it’s part of an ongoing retraining or updating mechanism, which would increase costs.

Option D: Involves an agent and a fine-tuned LLM, which could be overkill and lead to higher development and operational costs.

Option Ais the most suitable as it provides a cost-effective, minimal development approach while ensuring the application remains up-to-date with new information.

The problem describes a Retrieval-Augmented Generation (RAG) system for HR policy Q&A where responses are incomplete and unstructured due to the retriever failing to return sufficient context. The engineer has already tried different embedding and response-generating LLMs without success, suggesting the issue lies in the retrieval process—specifically, how documents are chunked and indexed. Let’s evaluate the options.

Option A: Add the section header as a prefix to chunks

Adding section headers provides additional context to each chunk, helping the retriever understand the chunk’s relevance within the document structure (e.g., “Leave Policy: Annual Leave” vs. just “Annual Leave”). This can improve retrieval precision for structured HR policies.

Databricks Reference:"Metadata, such as section headers, can be appended to chunks to enhance retrieval accuracy in RAG systems"("Databricks Generative AI Cookbook," 2023).

Option B: Increase the document chunk size

Larger chunks include more context per retrieval, reducing the chance of missing relevant information split across smaller chunks. For structured HR policies, this can ensure entire sections or rules are retrieved together.

Databricks Reference:"Increasing chunk size can improve context completeness, though it may trade off with retrieval specificity"("Building LLM Applications with Databricks").

Option C: Split the document by sentence

Splitting by sentence creates very small chunks, which could exacerbate the problem by fragmenting context further. This is likely why the current system fails—it retrieves incomplete snippets rather than cohesive policy sections.

Databricks Reference: No specific extract opposes this, but the emphasis on context completeness in RAG suggests smaller chunks worsen incomplete responses.

Option D: Use a larger embedding model

A larger embedding model might improve vector quality, but the question states that experimenting with different embedding models didn’t help. This suggests the issue isn’t embedding quality but rather chunking/retrieval strategy.

Databricks Reference: Embedding models are critical, but not the focus when retrieval context is the bottleneck.

Option E: Fine tune the response generation model

Fine-tuning the LLM could improve response coherence, but if the retriever doesn’t provide complete context, the LLM can’t generate full answers. The root issue is retrieval, not generation.

Databricks Reference: Fine-tuning is recommended for domain-specific generation, not retrieval fixes ("Generative AI Engineer Guide").

Conclusion: Options A and B address the retrieval issue directly by enhancing chunk context—either through metadata (A) or size (B)—aligning with Databricks’ RAG optimization strategies. C would worsen the problem, while D and E don’t target the root cause given prior experimentation.