What a RAG knowledge assistant costs to build, partner or buy
For a leader weighing an internal document assistant against SaaS and a partner, the cost depends on document volume, permission complexity, model hosting and how much evaluation the answers require. The open-source engines are usually free to download, but labor, infrastructure and upkeep are not. Those costs fall on your team or a partner, they recur, and they grow with the number of documents, users and models you run. Two license terms can change the answer before any labor is priced. Dify's license restricts operating a multi-tenant service from its source code without written authorization, and its self-hosted Community edition is limited to a single workspace, so a multi-team or customer-facing build may need the paid Enterprise edition, as described on the Dify pricing page. Open WebUI's license prohibits altering or removing its branding in deployments above 50 users in any rolling 30-day period unless you hold an enterprise license or written permission, as set out in the Open WebUI License. Confirm both terms against your user count and tenancy model with counsel before you select a platform.
Ongoing costs are the ones most often missed. Ingestion and parsing pipelines must be rerun when source documents change, the vector database needs block storage with SSD or NVMe drives and, for production, at least three nodes so quorum survives a failure, and embedding and inference compute recur with every query and re-index. Evaluation work, including a test set of questions with expected sources, is a repeated effort. Ask any partner for a quote that separates one-time build, monthly infrastructure, model usage and evaluation upkeep, and that states which of these DeployOpen would deliver under a fixed written scope. DeployOpen does not publish a price list, so the quote follows from your agreed scope.
Which costs sit in-house, with a partner or in SaaS
This table compares who pays for each part of a RAG knowledge assistant under three options: building in-house on self-hosted open source, engaging an implementation partner, or buying a SaaS or hosted RAG product. Read each row for the payer and whether the cost is one-time or recurring, then confirm the figures in written quotes. The main takeaway is that open-source licenses remove license fees for the core software, but infrastructure, evaluation and upkeep stay with whoever operates the stack.
| Cost component | Build in-house (self-hosted) | Implementation partner | SaaS or hosted RAG |
|---|---|---|---|
| Ingestion and parsing pipelines | Your engineers build and maintain connectors and parsing jobs; recurring as sources change | Partner scopes, builds and hands over connectors; one-time build, then your upkeep unless retained | Often bundled in the subscription; check connector coverage and scanned-document handling |
| Vector database | Self-hosted Qdrant is Apache-2.0 with no license fee; you pay for servers, storage and operations | Partner sizes and deploys on your infrastructure; you still pay hosting and storage recurrently | Qdrant Cloud free single-node tier exists; Standard is usage-priced with a 99.5% uptime SLA; Premium has a minimum spend |
| Embedding and inference compute | You pay model hosting or API usage; recurring and driven by document volume and query load | Same compute costs, plus a one-time sizing and tuning effort | Usually metered per user or per usage; Glean draws Model Hub usage against an upfront commit |
| Application layer license | Open WebUI and AnythingLLM are free to self-host under their terms; Open WebUI branding rules apply above 50 users in any rolling 30 days | Enterprise terms are quoted; confirm license scope for your user count | Dify Cloud Professional is listed at 590 US dollars per year and Team at 1,590 US dollars per year (Dify pricing page) |
| Dify self-hosted edition | Community edition is free and limited to one workspace; SSO and multiple workspaces need Enterprise | Partner can quote Enterprise, which is custom priced and includes SSO and professional support | Not applicable; the Dify Cloud plans listed above are hosted by Dify |
| Evaluation and testing | Your team builds the test set, runs retrieval and answer checks, and repeats them after each change; recurring | One-time evaluation design plus optional retained testing; confirm the scope in writing | Usually limited to vendor dashboards; confirm whether you can run your own answer evaluations |
| Observability and logging | Self-hosted tracing and log storage cost you storage and staff time; recurring | Partner configures logging to meet your retention needs; you keep the storage cost | Vendor-held logs; confirm retention, export rights and fit with your record-keeping duties |
| Identity integration | Your team configures OIDC, LDAP or SCIM and permission mapping; one-time plus ongoing review | One-time integration and permission testing, with support available under a retainer | Often sold as a higher tier; confirm SSO and permission-aware retrieval are included |
| Maintenance and upgrades | Your team patches, upgrades and backs up every component; recurring staff time | Upgrades and support run under a retainer or a fixed-scope agreement; confirm the terms | Vendor applies upgrades; you pay the subscription and accept the vendor release cadence |
What moves the cost of a RAG assistant up or down
Most of the cost of a RAG knowledge assistant comes from the corpus, the access rules and the operating model around the vector store, not from the software license. The drivers below connect directly to the build, partner or SaaS decision, and each one is something to scope in writing before a quote is compared.
Document count and file types
The number of documents, their formats and how many are scanned images or complex tables set the size of the ingestion and parsing pipeline. Clean text is cheap to index, while scanned manuals need OCR, layout parsing and re-checking of chunks, which adds effort before any model runs. Ask the partner to inventory sources by file type and page count, and to state which parsing steps are included in the fixed scope and which are priced as extras.
Permission mapping from source systems
Carrying source-document permissions into answers means mapping access rules from file shares, wikis and identity groups into retrieval filters, then testing that restricted content never reaches unauthorized users. This is one of the largest one-time costs and it recurs whenever group membership or folder rules change. Ask how permissions are synced, how often, and who tests the filters.
User count and Open WebUI branding
Open WebUI's license clause prohibits altering or removing Open WebUI branding in deployments above 50 users in any rolling 30-day period, unless the operator holds an enterprise license or written contributor permission. Above that threshold, white-labeling becomes a license line item, with pricing quoted by sales after an expected seat count is given. Confirm your peak active user count with counsel and the vendor before choosing the interface.
Vector database operating model
A single Qdrant node is simple to run, but it is a single point of failure and it does not support the zero-downtime upgrades that the enterprise operator offers. Distributed mode with at least three nodes adds Raft-based coordination and a replication factor that stores extra copies, and the replication factor of 2 doubles storage. Storage must be block-level SSD or NVMe, not NFS or S3, which affects the infrastructure bill.
Model hosting and GPU sizing
Hosted model APIs bill per token, so cost follows query volume and document context length. Self-hosted inference on Ollama or vLLM moves cost into GPUs or servers, a recurring infrastructure charge that must be sized for peak concurrency and the model chosen. Ask the partner for a sizing model based on expected queries per day, average context size and latency targets, and confirm which GPU types the chosen model needs.
Evaluation depth and maintenance
A retrieval test set with known answers, citation checks and regression runs after each re-index or model change is what keeps answer quality measurable. Its cost scales with the number of document domains, the review hours from subject matter experts, and how often the platform is upgraded, since Dify minor releases arrive roughly monthly. Ask whether evaluation is a one-time deliverable or a retainer, and who owns the test set after handover.
Which RAG costs are one-time and which recur every month
The one-time cost of a self-hosted RAG knowledge assistant is mostly design and setup work: architecture, connector builds for each document source, the initial ingestion and parsing run, mapping source permissions into the retrieval layer, building the evaluation set of questions with expected sources, and hardening the deployment. Much of this is paid to whoever does the build, whether an internal team or a partner under a fixed-scope agreement, so the quote should itemize each item separately. Ask for the connector list by source system, how permission inheritance is tested, and who writes and signs off the evaluation set. Note that the evaluation set is one-time in creation but recurs whenever the model, chunking or embedding changes, so it should be budgeted as both. The vector database and the chat front end also need a one-time configuration pass, including **TLS and API key settings** on Qdrant, which are not enabled by default in a self-deployed instance.
The recurring cost is the larger share over the life of the assistant, and it comes from keeping the stack current. Dify minor releases arrived roughly monthly in the period covered by the researched facts (1.16.0 on July 17, 2026, 1.17.0 on August 25, 2026 and 1.17.1 on September 10, 2026), and Open WebUI published v0.12.3 on October 16, 2026. Each upgrade needs regression testing against the evaluation set, so upkeep is a standing labor line. Other recurring items are re-indexing when source documents change or the embedding model is replaced, upgrades to the model runtime and the Qdrant engine, backup and snapshot jobs with tested restores, log retention for audit, and a support retainer if the organization does not staff these tasks internally. Ask each partner for the expected frequency of each item and the hours they assume, and confirm the upgrade and retainer terms in writing before you compare quotes.
Recurring costs also depend on infrastructure that the organization already pays for, and the quote should separate it from services. Compute for embeddings and inference, storage for vector indexes and object stores, and database hosting for PostgreSQL and Redis recur through your cloud or data center contracts. Qdrant documents that its distributed mode keeps extra copies through a replication factor, and a factor of 2 doubles storage, so replication choices directly raise the recurring storage bill. Dify Enterprise licensing and any Open WebUI enterprise license are separate recurring fees that are quoted, not published. Before comparing against a SaaS knowledge assistant, which charges per user, model the recurring fees on the same user count and document volume.
When building, partnering or buying SaaS makes sense
Building in-house means your engineers select, deploy and operate each component: the parsing and ingestion pipeline, the vector database, the embedding and chat models, the access layer and the evaluation harness. Everything runs in your accounts, and the cost is mostly staff time plus the compute you provision. Expect the largest share of cost to be labor for ingestion tuning, retrieval testing and ongoing re-indexing rather than software fees, since most of the core components are open source.
- Team skills: Platform engineers for Kubernetes or Docker operations, data engineers for document parsing and chunking, and someone who can run retrieval evaluations and interpret the results.
- Infrastructure control: Documents, embeddings and prompts stay in your environment, and you decide the upgrade and patch cadence.
- Vector database operations: Qdrant needs block-level storage and is not secure by default, so authentication, TLS and backups are your responsibility from day one.
- Licensing exposure: Components such as Open WebUI carry branding terms once you pass 50 users in any rolling 30-day period, and Dify's license restricts multi-tenant use, so check each license against your user counts.
- Ongoing load: Releases arrive frequently, so budget recurring time for upgrades and security reviews.
Before committing, ask who will own on-call duty for the assistant and how many hours per month your team can give to retrieval quality work.
A fixed-scope implementation partner plans, deploys, integrates, hardens and documents the assistant inside your infrastructure, then hands over the runbooks and the working system. The partner is paid for agreed deliverables, and recurring support is usually a separate retainer. Your data and models stay in your cloud accounts, data center or air-gapped network, so residency decisions remain yours. Cost is front-loaded into scope definition, connector work, evaluation design and handover, followed by a smaller recurring line for operations if you buy it.
- Team skills needed from you: A named owner who can accept the design, provide access and make decisions on document permissions, plus a platform team to receive the handover.
- What stays inside: Source documents, vector indexes, logs and model endpoints run in your infrastructure, and the partner works in that environment.
- Scope control: Agree sources, file types, user counts and evaluation criteria in writing before work starts, since these drive the quote.
- Licensing exposure: Commercial terms for components such as Dify Enterprise or Qdrant Hybrid Cloud or Private Cloud are priced through sales, so request those quotes as line items.
- Retainer exposure: Upgrades, continuous testing and support are separate agreements, so decide early whether you need them.
Before signing, ask the partner to show how the written scope defines acceptance tests for retrieval accuracy and permission handling.
A SaaS knowledge assistant or hosted RAG platform gives you a working assistant with connectors, hosting and vendor-managed upgrades for a subscription. Cost is primarily recurring, often charged per user or per usage tier, and it can rise with seat counts and model usage. The vendor hosts indexes and inference, so you accept their data residency, subprocessors and retention terms. Some vendors offer hybrid options, but the vendor still controls the platform roadmap and pricing.
- Team skills needed: A small administration group for connectors, access policies and user onboarding, with less need for vector database or model operations staff.
- Data residency trade-off: Documents and embeddings typically leave your environment, so confirm regions, retention and whether the vendor can restrict processing to your jurisdiction.
- Subscription exposure: Per-user or usage-based charges recur, and some features draw on separate usage commitments, so model costs at your expected peak.
- Lock-in: Export formats for indexes, prompts and evaluation data differ by vendor, so ask how you would migrate.
- Sovereignty limits: Regulated workloads may not meet your residency or audit requirements without contractual evidence.
Before buying, request the vendor's data processing terms, exit process and total recurring cost at your projected user count.
How to build your own RAG cost estimate
This sequence turns an unscoped idea into a cost model with named line items, so you can compare a self-built stack against a partner-delivered scope or a SaaS subscription on the same basis. Each step produces a figure or decision that feeds the final comparison.
Inventory document sources and volumes
Start by listing every repository the assistant should read, such as file shares, wikis, ticketing exports, policy portals and email archives. For each source, record the file types, page counts, share of scanned images, total storage size and how often content changes. Ingestion and parsing effort depends mostly on these figures, so they are the first input to request from your own IT teams and from any connector vendor. The step is done when you have one table with source, format, volume and change rate, reviewed by the content owners who will be accountable for accuracy.
Classify permissions on each source
Next, mark which sources carry access restrictions and how those restrictions are enforced today, through groups in your identity provider, folder permissions or document labels. Permission mapping adds integration work and, in some designs, a second retrieval filter, so it changes both build effort and the scope a partner must price. Ask your identity team to confirm group structures and whether any source needs per-document rules. The output is a permission matrix that identifies which collections must be separated before indexing.
Choose the vector database and deployment path
Select the vector store and decide where it runs. Qdrant's open-source engine is Apache-2.0 licensed and free to self-host, while its managed Qdrant Cloud tiers are priced by usage or by minimum spend, and Hybrid Cloud and Private Cloud are quoted through sales. Qdrant's production paths also differ in features: the community Helm chart omits enterprise capabilities such as zero-downtime upgrades. Confirm the storage class you can provide, since Qdrant requires block storage and does not run on NFS or S3. The output is a named vector database, a deployment path and a list of the paid tiers you would need.
Size compute for embeddings and inference
Estimate the compute needed to embed the corpus once, to re-embed changed documents and to answer questions at expected concurrency. Embedding cost is largely one-time per corpus plus recurring for updates, while inference recurs with every query and scales with user count and answer length. Dify's self-hosted Docker Compose stack needs at least 2 CPU cores and 4 GB of RAM before any model load, and Qdrant's distributed mode recommends at least three nodes with a replication factor that doubles storage at a factor of 2. Ask your platform team for GPU or API options and request a written model-hosting estimate from each provider you consider.
Price licenses for the chosen platform
Match each application layer to its license terms before you add up costs. Open WebUI is under a license that requires branding to stay unless a deployment has no more than 50 end users in any rolling 30-day period or holds an enterprise license, and enterprise pricing is quoted after contacting sales with an expected seat count. Dify's Community edition is free but limited to one workspace, Enterprise is custom priced, and its license restricts operating a multi-tenant service. Record each license cost, whether it recurs, and who issues the quote. The output is a license table that can be read alongside the infrastructure figures.
List evaluation and observability work
Plan the work that proves answers are accurate and cited. This includes building a question set with expected sources, scoring retrieval and answer quality on each release, and tracing prompts, sources and costs in production. Open source tools such as Langfuse (MIT, except its ee directories) and Arize Phoenix (Elastic License 2.0, which prohibits offering it to third parties as a hosted service) can host these records in your own storage. Evaluation is partly one-time setup and partly recurring whenever the corpus or model changes. Ask a partner to put test design, reviewer time and log retention in a written scope.
Add a recurring maintenance line
Finish with the ongoing costs that persist after launch: upgrades, security patches, backups and snapshots, re-indexing, on-call coverage and access reviews. Dify ships minor releases roughly monthly, and Open WebUI and Qdrant also release regularly, so each upgrade needs a test cycle. Logging for high-risk use under the EU AI Act must be kept at least six months under Article 19(1) where it applies, which adds storage. Record the expected hours and tooling for each item, label which ones are internal staff and which are partner retainers, and compare the total against SaaS subscription fees for the same user count.
Questions buyers ask about RAG assistant cost and effort
Does open source mean a RAG assistant costs nothing to run?
No. The software licenses for the core components are often free, but the work and infrastructure around them are not. You pay for compute to parse and embed documents, for storage and backups of the vector index, for inference (GPUs or servers for the model), for the people who design and test retrieval and answers, and for ongoing upgrades. Some components carry license conditions instead of fees. Dify's Community edition is free but limited to a single workspace, and its license restricts running a multi-tenant service from the source code without authorization. Open WebUI's branding clause applies above 50 users in any rolling 30-day period unless you hold an enterprise license. Confirm these conditions against your user counts before you budget.
How long does a RAG assistant build take, and can the timeline be fixed?
The duration depends on the number of document sources, the file types (including scanned and tabular content), the permission model and the evaluation work required before release. DeployOpen agrees scope, deliverables and timelines in writing before work starts, and most engagements are fixed-scope, so a fixed timeline is possible only once the written scope exists. Any timeline given before a scoping review is an estimate, not a commitment. Ask the partner to show which assumptions the timeline depends on, such as the number of connectors, the size of the corpus and how many evaluation rounds are included.
What does the client team need to provide for the build?
The client team usually provides several inputs that the build cannot do without. These typically include:
- Named owners for each document source and its access rules
- A current identity provider setup for single sign-on and group mapping
- Subject-matter reviewers who write and check the test questions and judge answer quality
- Infrastructure approval for compute, storage, networking and backups in the client's own cloud account, data center or air-gapped network
Ask the partner to list these inputs in the written scope, with the person or role responsible for each and the date each is needed.
Can our own engineers build and run a RAG assistant without a partner?
Yes, a capable internal team can build a RAG assistant with these platforms, and the licenses for RAGFlow, AnythingLLM and Open WebUI permit internal use. The harder parts are often the ones that do not show up in a demo: keeping source permissions in retrieval results, measuring answer accuracy against a test set, re-indexing when documents change, hardening the stack and keeping up with monthly minor releases. Dify, for example, shipped minor releases roughly monthly in 2026. Consider whether your team can staff these tasks continuously, and whether a partner should cover the initial build, hardening or ongoing support through a retainer.
What do Dify Enterprise or Open WebUI enterprise licenses add, and what do they cost?
Dify Enterprise adds single sign-on, multiple workspaces, advanced security, a commercial license and professional support, and it can be deployed on Kubernetes for high availability. Open WebUI's enterprise licenses allow white-labeling and custom branding, with dedicated support options. Neither vendor publishes a price in the supplied information. The Dify pricing page lists Dify Cloud Professional at 590 US dollars per year and Team at 1,590 US dollars per year, but those are hosted plans, not self-hosted Enterprise pricing. Open WebUI's pricing is quoted after you contact enterprise sales with an expected seat count. Ask each vendor for a written quote that states the seat or workspace count, the term and the support level.
How do Qdrant Cloud tiers compare with self-hosting the vector database?
Qdrant's open-source engine is free to self-host under Apache-2.0, so the cost shifts to infrastructure and operations. Qdrant Cloud offers a free single-node tier with 0.5 vCPU, 1 GB RAM and 4 GB disk, a usage-priced Standard tier with a 99.5% uptime SLA, and a Premium tier with a minimum spend that adds SSO, private VPC links and a 99.9% uptime SLA. Self-hosting needs block storage (not NFS or S3), SSD or NVMe drives, TLS and API keys configured, and at least three nodes for a production cluster. Compare the cloud tier's uptime and data-location terms with the internal cost of running that cluster, and ask for quotes for Hybrid Cloud or Private Cloud if you need managed operation inside your own infrastructure.
Which recurring costs should we budget for after launch?
Recurring costs fall into four groups: inference compute for every answer (GPUs or servers, or model API usage if you use a hosted model), storage and backups for the vector index and source files, platform maintenance (monthly minor releases, security patches and compatibility checks), and continuous evaluation as documents change. Qdrant replication doubles storage at a replication factor of 2, and Open WebUI needs PostgreSQL, Redis and shared file storage for multiple replicas. Ask the partner for a written support model that states what is covered per month, how re-indexing is handled, and what happens when a vendor releases a breaking change.
What to put in a RAG cost quote before you sign
A comparable written quote starts with a fixed list of sources and connectors, because every document repository, file share, wiki or helpdesk export adds parsing, deduplication and refresh work. The quote should name each source, its file types, whether scanned content needs OCR, and how often each collection is re-indexed. It should also state the permission design: how source-document access rights are read, mapped to users or groups, and enforced at retrieval time. A quote that says only 'connect the document estate' leaves the largest cost driver undefined. Ask the partner to write the permission model down, including what happens when a source system's rights change.
The second block covers vector database topology and backup. Ask whether the quote includes a single Qdrant node or a distributed setup, and whether that choice is sized for the number of collections and expected growth. The Qdrant distributed deployment docs recommend at least three nodes for production clusters, so quorum survives a node failure, and a replication factor doubles storage at a factor of 2. Also confirm the backup method, such as Qdrant snapshots, the storage location, restore testing, and whether keys, TLS and audit logging are configured. The quote should also list the evaluation method: the test question set, who owns it, how answer accuracy and citation correctness are scored, and what re-testing follows each model or index change.
The final block covers hardening, license review, handover and retainer terms. Ask for written confirmation of the license terms for the chosen front end, including Open WebUI branding limits above 50 users in a rolling 30-day period or Dify's multi-tenant restriction, and whether an enterprise license is priced separately. Handover should list documentation, runbooks, admin training and who owns model and connector upgrades after go-live. DeployOpen works to fixed-scope written agreements and does not publish a price list, so the quote follows from this scope. Any line that stays vague is a cost risk to negotiate before signing.
How DeployOpen scopes and delivers a fixed-scope RAG implementation
DeployOpen plans, implements, customizes, integrates, hardens, documents and supports self-hosted open-source platforms inside the client's own infrastructure. For a RAG knowledge assistant, that means the cost model is set in a written scope covering ingestion, the vector database, model access, evaluation and handover before any build work begins.
How we scope
- Inventory document sources, file types and scanned content that the assistant must index.
- Confirm whether source permissions must carry into answers, since this changes connector and retrieval work.
- Record in the written scope which cost components DeployOpen delivers and which the client buys or runs.
How we build
- Deploy the chosen framework such as RAGFlow, AnythingLLM, Dify or Open WebUI in the client's own cloud account or data center.
- Configure a self-hosted vector database, with access controls and storage sized to the document corpus.
- Build an evaluation set from real questions and test retrieval quality and citation accuracy before release.
How we hand over
- Document the architecture, the license terms that apply to the chosen platform, and the cost drivers behind the build.
- Provide runbooks for re-indexing, backups, upgrades and vector database snapshots.
- Hand over evaluation results and the recurring cost model so the client can compare build, partner and SaaS options.
How we support
- Offer ongoing support, upgrades and continuous testing as a retainer after handover.
- Review platform releases and license changes, and plan upgrades on the client's schedule.
- Confirm with the client which support, model and hosting costs continue after the initial scope closes.
