Works with the AI Agents you already use
Agent Knowledge Architect — Grounded AI Knowledge Systems
Design a trustworthy knowledge foundation for AI agents, from authoritative sources and ingestion to grounded retrieval, citations, freshness, and ongoing evaluation.
$9.99
Agent Knowledge Architect — Grounded AI Knowledge Systems
Example session with this skill installed
Design a source-grounded knowledge architecture for an internal support assistant. It must answer policy questions from 18,000 docs (PDF, HTML, and tables), with weekly updates, English and French, role-based access, citations, and a <2 second p95 retrieval target. We have no chosen vector database or embedding provider. Propose source authority, ingestion and parsing, metadata and chunking, hybrid retrieval, context assembly, freshness/conflict handling, access controls, evaluation, rollout phases, key tradeoffs, and open questions. Clearly label assumptions and do not invent vendor capabilities.
- Read your context and instructions
- Compiled the agent knowledge architect
Knowledge architecture proposal: internal support assistant
Executive summary
Build a source-traceable retrieval system that answers only from authorized, sufficiently current evidence. Start with a small representative corpus and retrieval evaluation before selecting a vector database or embedding provider. The <2 second p95 target is a design constraint to measure end to end; it cannot be guaranteed from the information supplied.
Requirements and assumptions
Supplied: 18,000 PDF, HTML, and table-containing documents; English and French; weekly updates; role-based access; citations; target retrieval p95 under 2 seconds.
Not supplied: source systems and owners, document-level ACL metadata, authoritative-source hierarchy, deletion rules, hosting/data residency, query volume, freshness SLA, answer-generation model, and budget. Treat these as discovery items.
Proposed flow
Source and authority registry: inventory each repository, owner, format, update mechanism, effective date, and access model. Define a per-topic source-of-record ranking with the policy owner. Preserve source URL or stable ID and version.
2.
Ingestion and parsing: use incremental weekly sync with checksums and version detection. Parse PDFs and HTML into structured blocks; extract tables as rows/cells with headers and source locations, not flattened prose alone. Route scans or low-confidence extraction to OCR and quality checks. Quarantine parse failures for review.
3.
Canonical records and metadata: retain raw source, normalized text, language, title, section, source ID, version, owner, effective/expiry dates, ACL, ingestion time, and parser confidence. Keep provenance pointers through every derived chunk.
4.
Chunking and indexing: split by headings and semantic boundaries, preserving table structure and neighboring context. Record token/character length and overlap policy. Use lexical retrieval for exact policy terms and identifiers plus semantic retrieval for paraphrases; filter by role and language before candidate content reaches the model.
5.
Retrieval and answer assembly: retrieve a bounded candidate set, rerank for question relevance and authority, then assemble only the minimum passages needed. Return citations with source title, section, version/date, and link or stable ID. If evidence is missing, stale, inaccessible, or conflicting, state that plainly and request clarification or escalation.
6.
Freshness and conflicts: process additions, updates, and deletions idempotently. Mark superseded versions; do not silently mix them with current policy. Compare conflicting claims using the agreed authority hierarchy and effective dates; surface unresolved conflicts.
Security and governance
Carry source ACLs into retrieval and enforce authorization before content enters prompts, not only at display time. Test cross-role isolation, deletion propagation, auditability, retention, and prompt-injection handling in retrieved documents. Minimize stored personal data and define owners for access reviews and source changes.
Evaluation and rollout
Create a reviewed bilingual question set with expected source citations, including exact lookup, paraphrase, table questions, stale-policy traps, conflicting sources, inaccessible documents, and unanswerable questions. Measure recall@k, citation correctness, grounded answer accuracy, abstention quality, ACL leakage (target zero), freshness lag, parse coverage, and p50/p95 latency. Establish baselines before tuning.
Phase 1: source/ACL inventory and 100–200 representative questions.
Phase 2: ingest a small authorized slice and validate parsing, citations, and access filters.
Phase 3: compare lexical, semantic, and hybrid retrieval on the test set; benchmark latency with realistic concurrency.
Phase 4: expand sources, add monitoring and refresh/deletion controls, then limited user rollout with feedback.
Decisions needed next
Confirm source systems and owners, ACL granularity, authoritative policy sources, language-specific content coverage, deletion/retention requirements, expected query volume, deployment constraints, answer model, and what the 2-second target includes. Choose infrastructure only after those constraints and benchmark results are known.
Connects securely to your tools. The creator never sees your data.
About this skill
Build knowledge systems agents can trust
Agent Knowledge Architect turns a business or product requirement into an end-to-end knowledge architecture for RAG applications, enterprise assistants, support agents, research systems, internal copilots, and AI products. It maps the lifecycle from source selection and authority through ingestion, parsing, normalization, chunking, metadata, indexing, retrieval, reranking, context assembly, citation, freshness, conflict resolution, evaluation, and governance.
The design goal is minimum sufficient, high-quality evidence for the task—not maximum retrieved text. The architecture preserves provenance through transformations, respects source access boundaries, labels freshness, surfaces conflicts, and separates retrieved material from verified truth.
What it helps design
- Source inventories, systems of record, ownership, and authority ranking
- Parsing strategies for documents, tables, structured data, and OCR where supported
- Metadata, semantic chunking, versioning, deletion, and multilingual content
- Keyword, vector, filtered, hybrid, and multi-hop retrieval with reranking
- Context assembly, citations, provenance, freshness, supersession, and conflict handling
- Access controls, tenant isolation, sensitive-data protections, and knowledge governance
- Representative evaluation sets, quality metrics, feedback loops, and refresh plans
Working approach
The skill first clarifies target questions and decisions, then inventories actual sources, access, scale, update cadence, and constraints. It makes recommendations conditional on the tools and evidence supplied. It does not assume connectors, embeddings, vector databases, browsing, code execution, or live access. It never invents source content, citations, integrations, certifications, or current facts.
Best results
Provide the use case, example questions, source types and owners, approximate corpus size, formats and languages, update cadence, authorization model, hosting constraints, latency and budget targets, and any evaluation examples. Redact secrets and personal data.
Recommendations depend on the sources, access controls, infrastructure, and requirements provided; no connector or platform capability is verified automatically. Retrieval quality, citation correctness, latency, language coverage, and security must be measured on representative data. Parsing quality varies by document, especially scanned PDFs and complex tables. Resolve source ownership, ACLs, retention, and data residency with the responsible teams before implementation.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before any payment.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed 1 month ago
Permissions
No external accounts or services are required by default. Read only documents the user explicitly supplies or authorizes when available; respect source access controls. Do not connect to repositories, upload content, or expose sensitive data unless explicitly approved. Recommendations are architecture guidance and must be validated in the target environment.