Area guide · 3 packages · Trace ingestion through retrieval
Knowledge, retrieval and evidence
Turn authorized source material into retrievable, cited context while preserving its source, scope and freshness.
What are you trying to do?
- What is a knowledge space versus a vector-store collection?
- How does ingestion become usable evidence?
- Which provider and access boundaries does a knowledge host supply?
Why start with these packages
Suggested entry point
@lssm-tech/lib.knowledgeUse ingestion, retrieval, query, access and knowledge-runtime composition primitives.
Suggested entry point
@lssm-tech/integration.provider-embeddingInspect the concrete embedding provider boundary selected by an ingestion host.
Suggested entry point
@lssm-tech/integration.provider-vector-storeInspect the concrete vector-store adapter used by indexing and retrieval.
Suggested entry point
@lssm-tech/lib.context-storageStore the selected pack or snapshot when the caller needs a durable context identity.
Practical steps
Step 01
Trace one source document
Identify its knowledge space, namespace, source policy and original evidence location. Follow processing and indexing into a retrieval result, then check whether the answer preserves a useful citation back to the selected source.
Step 02
Compare retrieval modes
Use a static fixture to understand expected relevant passages and denied scope. Then inspect the equivalent vector-backed composition: provider selection, collection, ingestion revision and query behavior. Keep storage and model configuration explicit.
Step 03
Review a stale or removed source
Inspect the owning workflow’s behavior when source content changes, access is revoked or a required item is missing. The appropriate outcome may be reindexing, exclusion or an unavailable-context result rather than an apparently complete answer.
Choose the workflow, then the provider
The knowledge package contains static and vector retrieval, document processing, embeddings/indexing orchestration and retrieval-augmented queries. Its runtime convenience layer can compose ingest, retrieve and answer using supplied embedding, vector-store and LLM providers. Start with the space and namespace that express the intended source scope. Choose a static retriever for a deterministic source walkthrough or small fixture; choose vector retrieval when the application supplies a suitable indexed store and ingestion lifecycle. Neither choice removes access-policy responsibilities.
Keep the source and retrieval layers connected
The package exposes source-management operations and named knowledge spaces, including support FAQs, uploaded documents and Product Canon. Ingestion produces searchable material; retrieval selects material relevant to a query. Preserve original source identity, version and evidence locators through that transformation so a caller can inspect what supports an answer. Connector readiness manifests are useful for understanding supported composition paths, but the source explicitly distinguishes those manifests from live provider calls and production writes.
Bind access and context before answering
KnowledgeAccessGuard and grounding helpers provide the boundaries for scope-aware retrieval. The query result should remain connected to its allowed sources and citations. An agent context compiler can then enforce required inputs, trust classes and budgets. Avoid treating a retrieved passage as a new instruction authority or as permission to mutate a system. For knowledge mutations, the package separately documents dry runs, approval, idempotency, audit evidence and outbound-send gates; those checks belong to the selected mutation flow.
Follow a practical path.
Trace one source document
Identify its knowledge space, namespace, source policy and original evidence location. Follow processing and indexing into a retrieval result, then check whether the answer preserves a useful citation back to the selected source.
Compare retrieval modes
Use a static fixture to understand expected relevant passages and denied scope. Then inspect the equivalent vector-backed composition: provider selection, collection, ingestion revision and query behavior. Keep storage and model configuration explicit.
Review a stale or removed source
Inspect the owning workflow’s behavior when source content changes, access is revoked or a required item is missing. The appropriate outcome may be reindexing, exclusion or an unavailable-context result rather than an apparently complete answer.
Packages in this area (3)
3 of 3 shown
- Open manual →
Embedding provider implementations for ContractSpec.
Inspect the concrete embedding provider boundary selected by an ingestion host.
- Open manual →
Vector store provider implementations for ContractSpec.
Inspect the concrete vector-store adapter used by indexing and retrieval.
- Open manual →
RAG and knowledge base primitives
Use ingestion, retrieval, query, access and knowledge-runtime composition primitives.
What connects to this area
Schemas and results
Source records, queries and evidence references must keep their declared shapes.
Identity and access
Knowledge selection must use trusted tenant, actor and purpose context.
Agents and context
Retrieved content becomes classified input to a bounded agent context.
Intelligence and research
Research runtimes supply versioned evidence and suppression policy.