diff --git a/assert/architecture.svg b/assert/architecture.svg index 08a22d01..10f74ad4 100644 --- a/assert/architecture.svg +++ b/assert/architecture.svg @@ -1,22 +1,29 @@ - + Belief Context Graph architecture The Agent context manager streams evicted turns through the Graph HTTP service, while the Python SDK calls the same construction backends directly. Unified and hybrid backends share an incremental belief extraction, merge, relation linking, deterministic confidence, and persistence flow. Graph snapshots return to the Agent system prompt. - - + + - - + + - - + + - - + + + + + + + + + @@ -26,159 +33,160 @@ - + - + AGENT RUNTIME - - - + + + - - - Any Agent - Reference CLI included - User input · sessions · tools - - agent-agnostic - - - BCG Context Manager - Initial user input is always retained - Latest N completed turns remain raw - Evicted turns → Graph · Graph → system prompt - - BCG mode - - - - Model + Tools - OpenAI-compatible inference - Search · shell · files · extensions - Assistant and tool turns return to context - - - + + + Any Agent + Reference CLI included + User input · sessions · tools + + agent-agnostic + + + BCG Context Manager + Initial user input is always retained + Latest N completed turns remain raw + Evicted turns → Graph + + BCG mode + + + + Model + Tools + OpenAI-compatible inference + Search · shell · files · extensions + Assistant and tool turns return to context + + + - INTEGRATION PATHS + INTEGRATION PATHS - - - + + + - Graph Construction HTTP Service - One ordered streaming session per problem_id - Threaded across independent trajectories + Graph Construction HTTP Service + One ordered streaming session per problem_id + Threaded across independent trajectories - - - - - /turns - /graph - /finalize - /release + + + + + /turns + /graph + /finalize + /release - Python SDK - BCGRunner — run and session orchestration - BCGMemory — observe, read, search, context - Direct backend access without the HTTP layer + Python SDK + BCGRunner — run and session orchestration + BCGMemory — observe, read, search, context + Direct backend access without the HTTP layer - Lazy Backend Registry - shared contracts · config · sessions + Lazy Backend Registry + shared contracts · config · sessions - - - evicted completed turns - - + + + + evicted completed turns + + - CONSTRUCTION BACKENDS + CONSTRUCTION BACKENDS - - + + - - UNIFIED - General-purpose Graph LLM - Node, entity and stance extraction · bounded relation windows - Single configurable OpenAI-compatible model path + + UNIFIED + General-purpose Graph LLM + Node, entity and stance extraction · bounded relation windows + Single configurable OpenAI-compatible model path - - HYBRID - Graph LLM + local models - Semantic chunks · embeddings · DeBERTa stance · spaCy NER - Separate extraction and relation generation components + + HYBRID + Graph LLM + local models + Semantic chunks · embeddings · DeBERTa stance · spaCy NER + Separate extraction and relation generation components - - + + - - SHARED INCREMENTAL GRAPH SEMANTICS + + SHARED INCREMENTAL GRAPH SEMANTICS - - - - - - - - 1 · Split / chunk - role-tagged turn - 2 · Extract - beliefs + evidence - 3 · Merge - canonical nodes - 4 · Link - typed relations - 5 · Confidence - deterministic posterior - 6 · Snapshot - graph + artifacts - - - - - - + + + + + + + + 1 · Split / chunk + role-tagged turn + 2 · Extract + beliefs + evidence + 3 · Merge + canonical nodes + 4 · Link + typed relations + 5 · Confidence + deterministic posterior + 6 · Snapshot + graph + artifacts + + + + + + - - + + - - - GRAPH SNAPSHOT - beliefs · relations · confidence - + + + GRAPH SNAPSHOT + beliefs · relations · confidence + diff --git a/assert/benchmark_overview.svg b/assert/benchmark_overview.svg index 3d32b659..483af49d 100644 --- a/assert/benchmark_overview.svg +++ b/assert/benchmark_overview.svg @@ -1,5 +1,5 @@ - Efficient Context · Better Performance · Belief-Aware Reasoning + Less Tokens · Better Performance · Belief-Aware Reasoning Full-dataset Default, BCG, and Summary accuracy and mean token cost, followed by Default and BCG cumulative token cost by Agent-step depth for BrowseComp and BrowseComp-ZH. @@ -8,7 +8,7 @@ - Efficient Context · Better Performance · Belief-Aware Reasoning + Less Tokens · Better Performance · Belief-Aware Reasoning Default diff --git a/assert/benchmark_summary.svg b/assert/benchmark_summary.svg index c8c53db5..c56b68a3 100644 --- a/assert/benchmark_summary.svg +++ b/assert/benchmark_summary.svg @@ -1,10 +1,10 @@ -Efficient Context · Better Performance · Belief-Aware Reasoning +Less Tokens · Better Performance · Belief-Aware Reasoning Full-dataset Default, Summary, and BCG accuracy and mean token cost for BrowseComp and BrowseComp-ZH, shown with non-zero y-axis baselines. The light Summary segment is Summary Generation, and the light BCG segment is Graph Construction. -Efficient Context · Better Performance · Belief-Aware Reasoning +Less Tokens · Better Performance · Belief-Aware Reasoning Default Summary Summary Generation diff --git a/scripts/plot_benchmark_overview.py b/scripts/plot_benchmark_overview.py index 7377aa7b..5fb41bcf 100644 --- a/scripts/plot_benchmark_overview.py +++ b/scripts/plot_benchmark_overview.py @@ -38,7 +38,7 @@ def compose(summary: Path, horizon: Path, output: Path) -> None: }, ) title = ET.SubElement(root, _tag("title"), {"id": "title"}) - title.text = "Efficient Context · Better Performance · Belief-Aware Reasoning" + title.text = "Less Tokens · Better Performance · Belief-Aware Reasoning" description = ET.SubElement(root, _tag("desc"), {"id": "desc"}) description.text = ( "Full-dataset Default, BCG, and Summary accuracy and mean token cost, " diff --git a/scripts/plot_benchmark_summary.py b/scripts/plot_benchmark_summary.py index d10ec2dc..50802426 100644 --- a/scripts/plot_benchmark_summary.py +++ b/scripts/plot_benchmark_summary.py @@ -134,12 +134,12 @@ def render_svg(output: Path) -> None: token_centers = (858.25, 1072.75) parts = [ f'', - 'Efficient Context · Better Performance · Belief-Aware Reasoning', + 'Less Tokens · Better Performance · Belief-Aware Reasoning', 'Full-dataset Default, Summary, and BCG accuracy and mean token cost for BrowseComp and BrowseComp-ZH, shown with non-zero y-axis baselines. The light Summary segment is Summary Generation, and the light BCG segment is Graph Construction.', '', '', f'', - f'Efficient Context · Better Performance · Belief-Aware Reasoning', + f'Less Tokens · Better Performance · Belief-Aware Reasoning', f'Default', f'Summary', f'Summary Generation',