Retrieval-augmented generation built around your client's content — with a chunking strategy, reranking, real citations, and an honest "I don't know" path when the corpus doesn't have the answer.
Retrieval pod, predictable spine
These are the conversations agency owners describe when a knowledge-grounded brief is in front of them and the in-house team has wired up an embedding API but not yet a retrieval system that survives a real audit.
The chatbot looks helpful and is making things up — wrong policy, wrong price, wrong contact. We re-architect the surface around real retrieval, citation rendering, and an evaluated "I don't know" path so confidence and accuracy line up.
Half the questions are answered in one source and half in another. We build the ingestion, normalisation, and chunking layer once, with per-source tuning, so the retriever sees the corpus the way the buyer thinks about it.
Top-k chosen by feel, no reranker, no eval set, no failure mode for empty retrieval. We add hybrid search, a reranker, and a labelled question set so retrieval quality is measured, not vibed.
Compliance, legal, regulated content — the brief asks for traceable answers, not opaque ones. We make citation discipline a first-class part of the system: every claim ties back to a chunk a user can open, with run-time logs the buyer's auditor can read.
Retrieval work runs eval-first. The pod doesn't ship a grounded answer surface into a real workflow without a labelled question set and a documented citation contract behind it.
Working session with your delivery lead and the buyer-side stakeholder. We map the corpus, the expected query types, the citation contract, and what a "wrong" answer looks like. NDA and SOW signed under Salt Technologies, Inc.
Architecture readout: chunking strategy per source, embedding model, vector store, hybrid search and reranker, citation rendering, the "I don't know" path. In parallel, we build the labelled question set every retrieval change runs against in CI.
Iterative build with weekly working review. Ingestion pipeline first, then embeddings and the vector store, then retrieval and reranking. Eval pass rate, citation accuracy, and latency tracked from build one — not retrofitted at the end.
Production deployment behind a feature flag, with eval gates on every release, observability for retrieval quality and cost, and a written runbook for stale-document handling and re-indexing. Post-launch transitions cleanly into a Support & Maintenance retainer for corpus drift, model upgrades, and v1.x feature work.
Retrieval work is long-arc once a real corpus shows up. The two shapes below are the engagements where production-grade RAG systems actually land — a new service line under your brand, or a delivery pod inside an active client SOW.
Capability lead + delivery support
Standing RAG & Knowledge AI up as a defensible new service line on your rate card — capability lead, pricing template, monthly retainer with quarterly review.
Retrieval pod inside your SOW
Active client work that has been sold and now needs end-to-end RAG throughput inside your shared channel — long-arc, structured, named delivery owner.
Starting weekly rate for a single-capability AI pod. Multi-track AI pods (RAG + Agents, RAG + Integrations) and pods that mix capabilities are quoted at the highest applicable rate. Final SOW is scoped against the brief.
We work inside the retrieval tooling your team and your client already use — no parallel platform, no "we'll just rebuild it our way" surprise.
RAG & Knowledge AI shows up most often where the agency is already shipping AI work and the next brief asks for grounded, source-backed answers — not just a smarter chat box.
HubSpot partners shipping knowledge portals, content-grounded chatbots, and support-deflection surfaces over the HubSpot CMS, blog, and help content — where the standard search ends and a real retrieval system begins.
Boutique design and product studios shipping client-facing knowledge surfaces — brand-grounded answer experiences, internal portals, customer-content discovery — where retrieval quality is part of the design, not an afterthought.
Shopify partners adding merchant docs retrieval, product-knowledge assistants over catalogs, and shopper-side answer surfaces grounded in the actual store content — not in the model's guess at it.
Every RAG & Knowledge AI engagement runs on the same operating spine that protects long-arc retainers and Dedicated Partner Pods — contracted through Salt Technologies, Inc.
We don't email your client, join their calls, or appear in the proposal — unless you explicitly white-list a named engineer in the SOW.
We work in your GitHub, your model-provider accounts, your hosting, and your shared channel under aliases that fit your team's naming.
Mutual non-solicitation written into every MSA, with a defined window after the engagement ends. Same clause across every track.
MSA, NDA, and engagement SOW are issued by Salt — the Delaware C-Corp behind AgencyRelay.
The same operating spine sits underneath every AgencyRelay capability. Read the no-poach and confidentiality page for the contractual instruments behind these defaults.
Most agency owners deciding between a retrieval pod, a single AI specialist hire, or a freelance prompt engineer weigh the same trade-offs. These pages address them head-on.
All four AI capabilities side by side — useful when the brief sits between RAG and one of the other tracks and you want to see the routing decision laid out.
The most common shape for adding RAG & Knowledge AI to your rate card — capability lead, pricing template, quarterly review on whether the new line earns its place.
When a retrieval pod is cheaper, faster, and lower-risk than the next senior AI hire — and when staffing up is the right call.
Direct answers to the questions that come up on almost every RAG & Knowledge AI scoping call.
RAG grounds an answer in your client's content — the model retrieves and cites, but doesn't act. Agents take actions — they call tools, update systems of record, kick off workflows. Most agentic systems we ship combine both (an agent with a RAG tool in its toolbox), but the *primary* capability on the SOW is the one that defines the system. We make this call inside the brief & feasibility step in week one.
Three layers. First, a citation contract — the model is constrained to answer from retrieved chunks and surface the source for each claim. Second, an evaluated "I don't know" path that fires when retrieval is empty or low-confidence, instead of letting the model reach for its training data. Third, an eval suite that runs in CI on every retrieval or prompt change against a labelled question set the buyer signs off on, including hallucination scoring. All three are part of default scope.
We don't have a single default. The choice between pgvector (when the corpus is small-to-medium and Postgres is already in the stack) and a managed vector store like Pinecone, Weaviate, or Qdrant (for scale, hybrid filters, or multi-tenant routing) sits on corpus size, latency budget, and your client's existing data layer. Embedding choice — OpenAI text-embedding-3, Voyage, Cohere — is selected in the architecture readout with the trade-offs written down so your team and your client can refer back.
The corpus and the ingestion pipeline belong to your client. We build the ingestion, normalisation, and chunking layer once, with per-source tuning, and ship a re-index runbook so your client's team (or a Support & Maintenance retainer crew) can keep it fresh. Source-of-truth lives in your client's existing systems — Notion, Drive, Confluence, CMS, ticketing — and the vector store is a derived index, not a parallel content system.
Two patterns, picked in the architecture readout. For corpora that change daily — product docs, support content, CMS — we wire incremental re-indexing on source events (webhooks, polling, or change feeds). For slower-moving corpora — policy, legal, archived knowledge — we ship a scheduled re-index runbook with a freshness dashboard so stale chunks are visible before they become wrong answers. Either way, citation rendering shows the source's last-updated timestamp so users see freshness alongside relevance.
Production RAG systems aren't a one-week capability. The most common starting shape is a 4–6 week pod inside an Invisible Delivery Team SOW, sized around the corpus surface and the citation contract. For shorter retrieval work (a tightly scoped feature, an eval-suite stand-up, a chunking-strategy audit), we'll quote a tighter window against the same weekly rate.
The starting weekly rate for a single-capability AI pod is $3,200 per week. Multi-track AI pods (RAG + Agents, RAG + Integrations) and pods that mix capabilities (RAG + UI/UX, RAG + Backend) are quoted at the highest applicable rate. Final SOW is scoped against the brief; the rate is the floor, not a ceiling.
Most RAG systems graduate cleanly into a Support & Maintenance retainer post-launch — corpus drift monitoring, eval regression on new model versions, reranker re-tuning, citation-accuracy review, and v1.x feature work inside a monthly envelope. Either the same pod or a smaller maintenance crew carries it on the same MSA, no second sales cycle.
Yes. IP ownership and assignment on delivered code, prompts, evals, ingestion pipelines, retrieval logic, and supporting artefacts is written into the MSA — the work belongs to your agency (and onwards to your client per your own client contract) on payment of the relevant invoice. The Salt Technologies templates are counsel-reviewed and shared before signing.
Knowledge portal, support deflection, internal search, brand-grounded answers, or a regulated-content surface. Either way, the conversation starts with the work — not with a deck.
