Research-backed calculator guide

AI Knowledge-Base Answer Cost: Price Retrieval, Maintenance and Unanswered Queries

A practical method to estimate AI knowledge-base cost per answer, including model tokens, retrieval, reranking, maintenance, platform fees, and unanswered support work.

An AI knowledge base does not have one universal price per answer. The useful number is an allocated unit cost: the monthly cost of the model, retrieval layer, content upkeep, platform, and failed or escalated questions divided by the questions entering the system. If you count only output tokens, you will usually understate the cost of operating a support knowledge base.

What belongs in an AI knowledge-base answer cost?

Start with the unit boundary. “Answer” can mean a generated response, a question resolved without a human, or a support conversation closed by an agent. Choose one denominator and label it. I recommend tracking both incoming questions and successful self-service answers because they expose different problems.

1. Answer-generation tokens

The language model usually charges separately for input and output tokens. Input may include the user question, system instructions, conversation history, and retrieved passages. Output is the generated answer. Amazon Bedrock describes the same basic pattern: model pricing depends on the provider and model, and inference uses input and output token meters.^[1]

Do not use the visible answer length as the whole input. A concise 100-token answer may have been generated from a long system prompt and several retrieved chunks. Log the actual request and response usage whenever the provider returns it. If you are forecasting before launch, state an input-token assumption and a response-token assumption instead of presenting an estimate as an invoice amount.

2. Query embeddings and retrieval

A retrieval-augmented generation system may embed the question, search a vector index, optionally rerank the candidates, and then send selected context to the model. These steps are separate from generation. A managed vector provider may include some capacity in a plan and meter the rest. Pinecone, for example, lists database storage, read units, write units, inference, assistant tokens, and reranking as separate pricing dimensions.^[4]

Retrieval is not always a large per-question expense. It becomes material when the service has a minimum monthly commitment, when reranking is used on every query, or when the index has large storage and traffic requirements. Allocate the monthly minimum or fixed tier across the questions that actually use the knowledge base. If several products share one index, allocate by read units, query count, or another documented driver rather than assigning the entire bill to one workflow.

3. Ingestion, re-embedding, and maintenance

Initial indexing is a setup cost. Ongoing maintenance is an operating cost. New articles, changed policies, deleted documents, chunking, embedding refreshes, quality checks, and failed imports all consume time or provider usage. AWS’s worked chatbot example separates knowledge-base embeddings and storage from recurring query and response processing, which is a helpful structure for a total-cost model.^[2]

Track maintenance as its own line even if your model provider charges only fractions of a cent for embeddings. Editorial time is often the larger cost. A useful ledger has the article ID, owner, last verified date, change type, re-indexing tokens, and hours spent reviewing the resulting answers. That data also supports a knowledge base answer ROI calculation later.

4. Platform and fixed allocation

A vendor plan, vector database minimum, logging service, evaluation suite, or help-desk seat may not change when question volume changes. That does not make the cost irrelevant. Treat it as a fixed monthly amount and allocate it using the denominator that matches the decision.

For a new support bot, allocate platform cost per incoming question. For a product comparison, show the break-even volume: fixed monthly cost divided by variable contribution per question. Pinecone currently publishes a $20-per-month Builder plan and a $50 monthly minimum for Standard, while also listing usage prices beyond included allowances.^[4] Those figures are dated observations and can change, so use your contract or current provider page in a live budget.

5. Unanswered questions and escalations

A question that fails retrieval is not free. It may become a ticket, a live-chat handoff, a refund request, or a content-repair task. Record at least three outcomes: self-service answer, answer requiring human review, and no useful answer. Then attach a planning cost to the human path.

Do not call every escalation a model failure. Some questions should be routed to a person for safety, account access, policy exceptions, or missing transaction data. The cost model should therefore distinguish expected escalation from avoidable unanswered demand. The former is a design choice; the latter is a backlog opportunity.

Worked cost example

The following is an original planning example, not a quote from any provider. It uses 12,000 incoming questions per month and rounds only the final display. The model price assumption is OpenAI GPT-6 Luna at $0.10 per million input tokens and $0.50 per million output tokens as displayed on the checked pricing page.^[3] The Pinecone assumptions use its published $0.16 per million tokens for one listed embedding model and $2 per 1,000 rerank requests.^[4]

The direct model-and-retrieval rows total about $8.10, or $0.00068 per incoming question. That figure is useful for testing prompt length and model selection, but it is not the operating cost. After fixed platform allocation and maintenance, before escalation labor, the subtotal is $378.10, or $0.0315 per question. Including the assumed escalation work produces $1,242.10 ÷ 12,000 = $0.1035, rounded to $0.104 per incoming question.

There are 9,840 non-escalated questions in this example. If you divide the full monthly cost by those successful self-service outcomes, the result is $1,242.10 ÷ 9,840 = $0.126 per self-service answer. That number rises when the unresolved share rises, even if the model token rate stays constant. Conversely, cutting maintenance without measuring answer quality can make the unit cost look better while increasing stale or unsafe responses.

For a more complete budget, put the volume and token assumptions into the AI Agent Cost Calculator | Model Full Monthly Spend. If your main uncertainty is context size and retrieval, compare the assumptions with the RAG Cost Calculator: Estimate Retrieval, Embeddings and Generation. For token-only sensitivity checks, use the AI Token Cost Calculator: Input, Output and Cached Tokens Explained. When actual provider usage arrives, reconcile the estimate against invoice line items with the AI API Invoice Reconciliation Calculator.

How to make the estimate auditable

Create one row per cost driver and record whether it is a vendor price, a contract value, or an internal assumption. A good worksheet includes the date checked, currency, region, billing tier, included allowance, overage rule, and whether the number is recurring or one-time.

For token pricing, preserve input and output units separately. OpenAI’s pricing page shows that cached input can have a different rate from ordinary input and that Batch processing can be discounted relative to standard processing.^[3] If your application uses caching or batch ingestion, add those paths as separate rows rather than applying one blended rate to every request.

For retrieval, record the number of chunks searched, the number reranked, and whether long documents are split into multiple billable units. Cohere explains that a rerank search unit can cover up to 100 documents and that documents over the stated token limit can be split into multiple chunks for billing.^[5] That is why “one query” is not always one billable retrieval event.

For maintenance, sample a month of actual changes. Count new sources, changed sources, deleted sources, re-embedding runs, failed imports, evaluation runs, and human review hours. Then convert each to a cost. Do not hide a one-time migration inside the steady-state per-answer figure; show both launch cost and recurring cost.

What this estimate excludes

This example excludes the help-desk subscription, data storage outside the vector service, network egress, observability, evaluation-model calls, authentication, permissions engineering, compliance work, incident response, taxes, discounts, and regional price differences. It also excludes the business value of faster answers, the cost of an incorrect answer, and revenue or retention effects.

Those exclusions are deliberate. A calculator should make its boundary visible rather than imply that a low API bill equals a low total cost of ownership. AWS notes that model choice, embeddings, queries, responses, and concurrency affect a chatbot estimate.^[2] If concurrency requires reserved capacity, or if a provider minimum applies, replace the simple allocation with the relevant committed-capacity model.

Treat every number in this article as one of three types: a dated vendor-published observation, an arithmetic result from stated inputs, or an internal planning assumption. Vendor pages change. Your workload will differ. Re-run the worksheet when model, retrieval provider, document volume, answer policy, or escalation rate changes.

FAQ

What is the difference between cost per answer and cost per resolution?

Cost per answer divides spend by generated responses. Cost per resolution divides spend by conversations or issues that ended without further work. A knowledge base may generate an answer that does not solve the customer’s problem, so resolution is the stricter denominator. Track incoming questions, generated answers, and successful self-service resolutions separately.

Should maintenance be included in AI FAQ answer cost?

Yes, if the goal is an operating or unit-economics number. Include editorial and engineering time needed to review, update, re-chunk, re-embed, and evaluate the knowledge base. You can publish a direct AI-only number beside the fully loaded number, but do not present the former as the complete cost of serving answers.

How should I price an unanswered knowledge-base query?

Measure what happens next. If it becomes a human ticket, multiply the avoidable escalation count by an internal handling cost. If it triggers a content fix, record the repair time separately. If policy requires human review, classify it as expected escalation rather than as a retrieval defect. This lets you prioritize content gaps without pretending every handoff can be eliminated.

Does a vector database charge for every answer?

Not necessarily. Some plans include a free allowance or fixed capacity, while others meter storage, reads, writes, inference, reranking, or assistant tokens. Pinecone lists these as separate dimensions, and its illustrative examples explicitly exclude some inference and import costs.^[4] Check the provider’s unit definitions and allocate fixed minimums across the queries using the service.

Can a cheaper model reduce the cost per knowledge-base answer?

Usually it reduces the token component, but not automatically the total. A smaller model can require more context, retries, evaluations, or human review if answer quality falls. Compare fully loaded cost and successful-resolution rate, not only dollars per million tokens. Run the same evaluation set across candidate models and include the cost of the additional failure path.

What should I do when vendor pricing is not comparable?

Normalize the bill into meters: seats, questions, conversations, resolutions, tokens, storage, retrieval units, rerank units, and minimum commitments. Ask what counts as a resolution, whether an escalated thread is billable, what happens at the allowance limit, and whether ingestion is recurring. Keep quote-only or estimated values labeled as estimates; never convert an undisclosed quote into a precise per-answer claim.