RAG knowledge assistant — AI Stack

Components and roles

Every component earns its place.

01
Document parser
Extracts text and metadata while preserving source identity and access rules.
02
Embedding model
Maps passages and questions into vectors for candidate retrieval.
03
Qdrant reference store
Stores vectors, metadata, document versions, and permission filters.
04
Retrieval and answer service
Selects passages, builds the answer context, and returns citations.
05
Versioned evaluation set
Tests retrieval, answer support, refusal behavior, and freshness before release.

Data flow

Approved documents → parse and version → embed → Qdrant reference store → retrieve → answer with passages → evaluation and review

Default choices

Start here, then switch for a stated reason.

Vector store

Qdrant reference implementation

Selection reason: Provides explicit vector search, metadata, and filtering concepts for a portable reference design.
Switch when: Use pgvector when Postgres already owns operations; use managed search when hybrid retrieval and enterprise controls dominate.

Evaluation

Versioned task and passage-support set

Selection reason: Tests retrieval and answer support against the product’s actual corpus and questions.
Switch when: Add automated evaluation infrastructure when dataset size and release frequency make manual execution unreliable.

Operating assumptions

Cost, privacy, alternatives, and limits.

Cost assumption: Budget for parsing, embedding updates, vector storage, inference, evaluation runs, and maintenance—not only chat usage.
Privacy boundary: Document permissions must be enforced during ingestion and retrieval; the answer model must receive only authorized passages.
Reference implementation: Qdrant is a reference vector store, not a universal default. Use another store when existing infrastructure or filtering needs make it a better fit.
Alternatives: Use pgvector when Postgres is already the operational center; use managed search when hybrid keyword and vector retrieval or enterprise controls dominate.
Known limitation: Retrieval can miss relevant passages or surface stale content. Source citations and a passing demo do not prove complete or correct answers.

Human checkpoints

Keep accountable decisions visible.

Approve the source corpus and permissions
Review the evaluation set and failure cases
Own document freshness and incident response

Evidence and sources

See what supports this record.

Official info checked

Last reviewed 2026-09-05. Recheck current provider documentation, limits, pricing, and terms before adoption.

Every component earns its place.

Document parser

Extracts text and metadata while preserving source identity and access rules.

Embedding model

Maps passages and questions into vectors for candidate retrieval.

Qdrant reference store

Stores vectors, metadata, document versions, and permission filters.

Retrieval and answer service

Selects passages, builds the answer context, and returns citations.

Versioned evaluation set

Tests retrieval, answer support, refusal behavior, and freshness before release.

Data flow

Approved documents → parse and version → embed → Qdrant reference store → retrieve → answer with passages → evaluation and review

Start here, then switch for a stated reason.

Vector store

Qdrant reference implementation

Selection reason: Provides explicit vector search, metadata, and filtering concepts for a portable reference design.
Switch when: Use pgvector when Postgres already owns operations; use managed search when hybrid retrieval and enterprise controls dominate.

Evaluation

Versioned task and passage-support set

Selection reason: Tests retrieval and answer support against the product’s actual corpus and questions.
Switch when: Add automated evaluation infrastructure when dataset size and release frequency make manual execution unreliable.

Cost, privacy, alternatives, and limits.

Cost assumption

Budget for parsing, embedding updates, vector storage, inference, evaluation runs, and maintenance—not only chat usage.

Privacy boundary

Document permissions must be enforced during ingestion and retrieval; the answer model must receive only authorized passages.

Reference implementation

Qdrant is a reference vector store, not a universal default. Use another store when existing infrastructure or filtering needs make it a better fit.

Alternatives

Use pgvector when Postgres is already the operational center; use managed search when hybrid keyword and vector retrieval or enterprise controls dominate.

Known limitation

Retrieval can miss relevant passages or surface stale content. Source citations and a passing demo do not prove complete or correct answers.