Skip to content

architecture.md

Latest commit

 

History

History
116 lines (83 loc) · 6.88 KB

architecture.md

File metadata and controls

116 lines (83 loc) · 6.88 KB
Raw

Architecture Guidance

Deeper architectural detail lives in docs/architecture.md. This file covers non-obvious design decisions relevant to change-scoping.

Architectural Intent

contextweaver separates what to show the LLM (Context Engine) from which tools to offer (Routing Engine). These engines share types and stores but have intentionally different execution models:

  • Context Engine — async-first. Deals with I/O-bound operations (event log queries, artifact storage).
  • Routing Engine — sync-only. Pure computation (DAG traversal, beam search). No I/O.

This boundary is intentional. Do not propose making routing async "for consistency" — it adds complexity for zero benefit.

Non-Goals

contextweaver is not an LLM inference layer and not a tool execution runtime. It prepares context and routes tools but never calls models or executes tools. Feature proposals that cross these boundaries are out of scope.

Major Boundaries

Context Pipeline (8 stages)

The pipeline is a fixed sequence. Each stage has a single responsibility:

  1. generate_candidates — pulls events from stores into a candidate pool
  2. dependency_closure — ensures parent items (via parent_id) are included alongside their children
  3. sensitivity_filter — drops or redacts items above the sensitivity floor
  4. apply_firewall — summarises large outputs, stores raw data as artifacts
  5. score_candidates — ranks candidates by recency, tag match, kind priority, token cost
  6. deduplicate_candidates — removes near-duplicates via Jaccard similarity
  7. select_and_pack — greedily packs highest-scoring candidates into the phase token budget
  8. render_context — assembles the final prompt with BuildStats metadata

Why this order matters:

  • Dependency closure must happen before scoring, otherwise parents could be discarded before their children pull them in.
  • Sensitivity filtering before the firewall prevents sensitive data from reaching the summarizer.
  • Scoring after the firewall ensures scores reflect the summarised (not raw) content.

Routing Pipeline (4 stages)

  1. Catalog — register and manage SelectableItem objects
  2. TreeBuilder — convert flat items into a bounded ChoiceGraph DAG
  3. Router — beam-search over the graph for top-k relevant items
  4. ChoiceCards — render compact LLM-friendly cards (never full schemas)

Store Layer

All stores use typing.Protocol interfaces with in-memory defaults. This enables custom backends (database, Redis, etc.) without changing pipeline code.

  • EventLog — append-only. The audit trail.
  • ArtifactStore — raw tool outputs stored by the firewall. Supports drilldown via ViewRegistry.
  • EpisodicStore — short episodic memory entries.
  • FactStore — key-value facts persisted across turns.

Any backend can prove it honours these protocols with the shipped conformance kit (contextweaver.store.testing, issue #520): each check_*_conformance function takes a factory for an empty backend and asserts the round-trip, ordering, and not-found semantics the Context Engine relies on. For the ArtifactStore it also asserts that put() stamps a sha256 content_hash on the returned ref — the firewall's content-addressed idempotency short-circuit (#190) depends on it, so it is a protocol contract, not a backend detail. The bundled in-memory, JSON-file, and SQLite backends are all run through it in tests/test_store_conformance.py.

Thread-safety contract (issue #458)

The store protocols make no concurrency guarantee in their interface; each backend documents its own. The bundled backends:

  • InMemory* stores are not thread-safe. They are for single-threaded use and tests; guard them with your own lock for concurrent access.
  • JsonFileArtifactStore is single-process. Within one process it is thread-safe: put / delete / list_refs on a shared instance are serialised by an internal lock, and each individual file write is atomic (temp file + os.replace), so a reader never observes a torn or truncated artifact and a crash mid-write leaves the previous version intact. There is no cross-process advisory locking, so two processes writing the same base_dir are still unsupported.
  • SqliteEventLog opens its connection in WAL mode for single-process use; it is not shared across threads.

The gateway runtime (ProxyRuntime) inherits these guarantees through the store it is given: its read-only tool_view (drilldown) is safe to call concurrently against a JsonFileArtifactStore. A gateway that fans out to real concurrent clients should pick (or wrap) a backend whose contract matches its load — this is exactly what the protocol seam and conformance kit are for.

Sensitivity Enforcement

context/sensitivity.py is security-grade code. It enforces data classification (public → restricted) with two actions: drop or redact. The MaskRedactionHook is the built-in redactor. Changes to this module require extra review scrutiny — never weaken defaults.

Progressive Disclosure (ViewRegistry)

context/views.py provides a ViewRegistry that maps content-type patterns to view generators. When the firewall stores a large tool output as an artifact, the view system generates alternative representations (JSON subset, CSV summary, etc.) the agent can drilldown into without retrieving the full blob. drilldown_tool_spec() exposes drilldown as an agent-callable tool.

Key Tradeoffs

Decision Tradeoff Consequence of reversing
Protocol-based stores More files and indirection Allows backend swaps without pipeline changes
to_dict()/from_dict() + serde.py Two serialization paths Per-class methods handle class-specific logic; serde.py handles shared primitives. Consolidating loses encapsulation.
Sync routing / async context Two calling conventions Routing has no I/O — async would add overhead for zero benefit
8-stage pipeline Pipeline is long Each stage has a single well-defined responsibility. Merging stages creates coupling.
ChoiceCards never include schemas Limits LLM tool-call generation Keeps routing focused on which tool, not how — schema is provided at call-time via hydration

Structural Mental Model

Think of contextweaver as three layers:

  1. Data layer (types.py, envelope.py, config.py, serde.py, exceptions.py) — pure data, no I/O, no side effects.
  2. Store layer (store/, protocols.py) — stateful but simple append-only/read interfaces.
  3. Pipeline layer (context/, routing/, summarize/) — orchestration logic that reads from stores and produces output types.

Adapters (adapters/) convert external formats (MCP, FastMCP, A2A) into contextweaver types at the boundary.

Changes should flow within a layer. Cross-layer changes (e.g., adding I/O to the data layer) are red flags.