docs: RFC for adapting Agno's knowledge/vectordb architecture into HUF

[esafwan]

Comprehensive analysis and implementation plan for bringing Agno's vector database, embedder, reader, and chunker architecture into HUF's existing knowledge system.

Key points:

• Adapt interfaces (VectorDb, Embedder, ChunkingStrategy) not the orchestration layer — keep HUF's Frappe-native pipeline
• Phase 1: Embedder system via LiteLLM (25+ providers, zero new deps)
• Phase 2: 3 vector backends (Qdrant, ChromaDB, LanceDB)
• Phase 3: Enhanced readers (Excel, PPTX) and chunkers (recursive, markdown, code-aware)
• Phase 4: Hybrid search (BM25 + vector with RRF)
• Zero breaking changes to existing SQLite FTS5 users
• All new dependencies are optional extras
• Agno is Apache 2.0 licensed — compatible with adaptation

Relationship to scoped memory/data management

This PR should be read together with #274, which proposes the bridge between conversation data management, scoped Memory/Data Records, Knowledge Sources, Agno-style retrieval architecture, and Hindsight-style long-term memory.

In that combined model, this Agno-style architecture is the indexing/retrieval half of the bridge:

Memory/Data Record selected for knowledge
        ↓
Knowledge Projection
        ↓
Agno-style indexing pipeline
        ↓
FTS/vector/hybrid retrieval

This PR should not decide what becomes canonical memory. Instead, it should improve how selected/promoted records and normal Knowledge Sources are embedded, chunked, indexed, filtered, and retrieved.

Related:

• docs: add scoped memory and knowledge bridge RFC #274 — scoped memory/data management and knowledge bridge RFC
• Add Hindsight integration evaluation and recommendation (docs) #225 — Hindsight long-term memory evaluation

[esafwan]

TODO before this PR is taken up: update the RFC to reflect PR #280

Since this PR was opened, HUF adopted, via PR #280 (in review), the LlamaIndex-adapter pattern for vector backends and the get_advanced_config_schema() + generic advanced_config renderer for tuning params. This RFC should be updated to show which proposals are now implemented/superseded and which remain open.

• Update docs/rfcs/agno-knowledge-vectordb-adaptation.md to state that HUF uses thin LlamaIndex adapters (see pgvector_backend.py / chroma_backend.py in PR feat: add PGVector knowledge source #280) and HUF-side embeddings via huf.ai.knowledge.embedding.
• Mark the proposed separate Embedder class and huf/ai/knowledge/embedders/ module as superseded by the existing huf.ai.knowledge.embedding module.
• Mark the proposed Knowledge Source field changes (connection_config, search_type, vector_dimensions, etc.) as superseded by field groups + advanced_config; map any missing knobs to get_advanced_config_schema() entries.
• Mark Qdrant/Chroma/LanceDB backend ports as "future standardization PRs" that must follow the PR feat: add PGVector knowledge source #280 contract (subclass KnowledgeBackend, implement get_advanced_config_schema(), register via get_backend()).
• Keep the enhanced readers (Excel, PPTX) and chunkers (recursive, markdown, code-aware) as open proposals, since these are not covered by PR feat: add PGVector knowledge source #280 and can be pursued independently.
• Add a status section listing each proposal as Implemented / Superseded / Open, with links to PR feat: add PGVector knowledge source #280 and the cross-backend audit.

No code changes are required in this PR; once the RFC is updated it can be merged as documentation.