Your extraction pipeline processes invoices and extracts line items, subtotals, tax amounts, and grand totals. During evaluation, you discover that in 18% of extractions, the sum of extracted line item amounts doesn't match the extracted grand total—sometimes due to OCR errors in the source document, sometimes due to extraction mistakes by the model. Downstream accounting systems reject records with mismatched totals. What's the most effective approach to improve extraction reliability?
Answer(s): A
A: Add a "calculated total" field where the model sums extracted line items alongside a "stated_total" field. Flag records for human review when values differ. Correct.This preserves both sources of truth and enables reliable validation . Discrepancies can be flagged explicitly, improving accuracy without silently altering financial data.B: Extract line items and totals independently, then use a separate validation model to reconcile discrepancies by determining which extracted values are most likely correct. Incorrect.This adds complexity and uncertainty—“guessing” which value is correct can introduce errors in financial data.C: Add few-shot examples demonstrating invoices where extracted line items sum correctly to the stated total, encouraging the model to produce mathematically consistent extractions. Incorrect.Helpful but insufficient—does not handle OCR errors or real inconsistencies in source documents.D: Implement post-processing that automatically adjusts line item amounts proportionally when their sum doesn't match the stated total. Incorrect.This modifies financial data artificially, which is risky and unacceptable for accounting accuracy.
Your extraction system processes two document types: standard monthly reports (archived after processing) and urgent exception reports (must trigger business alerts within 30 minutes of receipt). Both use the same JSON schema. You want to minimize API costs while meeting latency requirements. How should you architect the processing pipeline?
Answer(s): D
A: Submit all documents to the Batch API with custom ids for tracking. When results arrive, immediately process urgent documents and trigger delayed alerts for exceptions. Incorrect.Batch processing introduces delays that can exceed the 30-minute requirement , making it unsuitable for urgent reports.B: Submit all documents to the real-time Messages API to ensure consistent processing latency across document types. Incorrect.This meets latency needs but is unnecessarily expensive for standard reports that don’t require real-time processing.C: Queue all documents and submit hourly batches, flagging urgent documents for expedited handling when batch results return. Incorrect.Hourly batching further increases delay and cannot meet urgent processing requirements.D: Route standard reports to the Batch API for 50% cost savings, and route urgent exception reports to the real-time Messages API. Correct.This balances cost efficiency and latency requirements , ensuring urgent reports are processed quickly while optimizing cost for non-urgent ones.
The extraction pipeline receives documents of varying types—some are invoices, others are contracts, and some are receipts. You've defined separate extraction tools, each with its own schema tailored to the document type. During testing, you observe that with tool_choice: "auto", Claude sometimes returns conversational text instead of calling an extraction tool, causing downstream parsing failures. You need guaranteed structured output without knowing the document type in advance. What's the most effective approach?
Answer(s): C
A: Consolidate all document types into a single unified-schema extraction tool and force that tool. Incorrect.This reduces extraction accuracy and creates an overly complex schema that doesn’t fit all document types well.B: Keep tool_choice: "auto" with system prompt instructions requiring tool use. Incorrect.Instructions alone are not enforceable, which is exactly why the issue occurs.C: Set tool_choice: "any" with all extraction tools defined. Correct.This guarantees that a tool will always be called , eliminating conversational text responses. While tool selection may not always be perfect, it ensures structured output every time , which is the primary requirement.D: Add a preliminary classification call, then make a second call with tool_choice forced to the identified extraction tool. Incorrect.While more precise, this adds extra latency and complexity. The question prioritizes guaranteed structured output , which C achieves more directly.
Monitoring shows 12% of extractions fall Pydantic validation with specific errors like "expected float for quantity, got '2 to 3". Retrying these requests without modification produces failures. What's the most effective approach to recover from these validation failures?
Answer(s): B
A: Set temperature to 0 to eliminate output variability and ensure consistent formatting Incorrect.This reduces randomness but won’t fix systematic extraction errors like misinterpreting ranges (“2 to 3”).B: Send a follow-up request including the validation error, asking the model to correct its output Correct.Providing specific validation feedback allows the model to correct the exact issue (e.g., convert “2 to 3” into a valid float), making recovery highly effective.C: Pre-process source documents to standardize problematic formats before sending them for extraction Incorrect.Helpful in some cases, but not scalable or sufficient for diverse real-world variations.D: Implement a secondary pipeline using a larger model tier to reprocess documents that fail validation Incorrect.More expensive and not necessary—targeted correction is more efficient and effective.
After three months of weekly sessions, your conversation history has grown to 85,000 tokens. When users ask "What did we conclude about the theme of isolation?", the assistant provides generic literary analysis rather than referencing the group's specific insights from earlier sessions. Discussions often build on previous meetings' conclusions, so maintaining narrative context is important. What's the most effective approach?
A: Add structured XML tags to mark significant discussion conclusions throughout the conversation history. Incorrect.Tagging helps organization, but it doesn’t solve the context length limitation or ensure relevant retrieval at query time.B: Use semantic embedding to index the full conversation history and retrieve only relevant past exchanges for each user query, replacing the linear conversation format with retrieved segments. Incorrect.Powerful for retrieval, but replacing the conversation entirely loses narrative continuity , which is important for ongoing discussions.C: Implement rolling window truncation to keep only the most recent 25,000 tokens. Incorrect.This discards earlier insights, which are critical for referencing past conclusions.D: Implement progressive summarization where older conversation blocks are replaced with concise summaries that explicitly extract key conclusions, decisions, and recurring themes, keeping recent exchanges verbatim. Correct.This preserves long-term context and key insights while staying within token limits, enabling the assistant to reference past conclusions effectively.
After a 40-minute session helping plan a dinner party, the conversation has grown to 78,000 tokens. The history includes: (1) the user mentioning a guest has a severe shellfish allergy, (2) measurements for scaling recipes to 8 servings, (3) the user's clarification that "room temperature butter" means 68°F in their kitchen, and (4) general back-and-forth about meal timing and presentation. You need to implement context management before the window limit is reached. What approach best balances information preservation with token reduction?
A: Summarize the entire conversation history into a concise summary capturing main topics discussed, then append new messages going forward. Incorrect.This risks losing critical details like allergies and precise measurements that must remain exact.B: Implement a sliding window retaining only the most recent 20,000 tokens relying on users to re-state important information when relevant. Incorrect.This discards essential context (e.g., allergy info), which is unsafe and unreliable.C: Store the full conversation externally and use semantic search to retrieve relevant portions for each turn, loading only matching segments into context. Incorrect.Helpful for retrieval, but may miss important persistent constraints (like allergies) if not explicitly retrieved every time.D: Extract critical structured data (allergies, serving counts, user-defined terms) into a compact reference section, summarize general discussion, and retain recent exchanges verbatim. Correct.This preserves high-priority facts exactly , reduces tokens efficiently, and maintains conversational continuity.
You're Implementing a feature where users refine their playlist preferences through multiple conversation turns. After deploying, you notice Claude's responses don't reflect what us earlier in the same conversation—for example, a user says they love jazz, but two messages later Claude asks what genres they enjoy. What is the most likely cause?
A: The model's context window has been exceeded by the conversation length Incorrect.This would only happen in very long conversations, and the issue appears early within just a few turns.B: The Claude API requires a session_id parameter that you haven't configured Incorrect.There is no required session_id—context must be explicitly managed by the application.C: Claude requires a vector database connection to maintain conversation memory Incorrect.A vector database is optional for retrieval, not required for basic conversation memory.D: Your application isn't including prior messages in the messages array Correct.Claude does not retain memory between calls—if previous messages aren’t included, it cannot recall earlier user preferences.
Your fitness coaching assistant uses a system prompt with detailed conditional logic: "If the user mentions being a beginner, provide step-by-step form instructions. If they use term 'progressive overload' or 'superset', respond concisely. If they ask about injury history, always recommend consulting a physician." During evaluation, you find the assistant correct explicit expertise declarations but struggles when users don't clearly state their level-often defaulting to overly detailed responses regardless of contextual cues like technical te Which change to the system prompt would most directly address this failure to pick up on implicit expertise signals?
A: Replace most conditionals with a general principle: "Adapt explanation depth to match user expertise, mirroring their terminology." Keep only the safety-critical conditional about consultations. Correct.This shifts from brittle rule-based logic to a flexible heuristic , enabling the model to infer expertise from subtle cues like terminology and tone.B: Add more conditional branches to cover additional expertise signals, such as "If user mentions specific rep ranges or asks about periodization, treat as advanced." Incorrect.This increases complexity and brittleness—edge cases will still be missed.C: Implement a pre-conversation intake that asks users to rate their experience level, then inject that rating into the system prompt as context for all subsequent responses. Incorrect.Useful, but doesn’t solve the issue of interpreting implicit signals during conversation .D: Add an explicit instruction for the model to ask a clarifying question about experience level whenever the user's expertise isn't immediately clear from their first message. Incorrect.This adds friction and interrupts flow instead of improving implicit understanding.
Post your Comments and Discuss Anthropic CCA-F exam prep with other Community members:
Our website is free, but we have to fight against AI bots and content theft. We're sorry for the inconvenience caused by these security measures. You can access the rest of the CCA-F content, but please register or login to continue.