diff --git a/README.md b/README.md index 3a56ef8e..552579eb 100644 --- a/README.md +++ b/README.md @@ -16,12 +16,23 @@ An agent should not have to reconstruct a project from scattered chat history on Engraphis turns local project knowledge into scoped, time-aware memory; retrieves the evidence that supports the current question; and returns a bounded, attributable context packet. +

+ Engraphis Knowledge Graph tab: force-directed entity-relation network +
+ Knowledge Graph · run engraphis-dashboard to see it live +

+

Diagram: project history becomes scoped and temporal Engraphis memory, hybrid recall, then bounded cited context for an agent
Store durable project knowledge · retrieve supporting evidence · give the agent only what it needs

+The flow is the essential path. See [measured context savings](#measured-quality-and-token-efficiency) +for reproducible fixture-level evidence of less returned content at the same tested retrieval +scores, without billing or latency claims. The sections below cover the dashboard, code graph, +local installation, governance controls, and hosted services in detail. + | Agent need | What Engraphis changes | |---|---| | Remember a project across sessions | Stores typed memory in a `workspace → repo → session` hierarchy and provides a last-session handoff. | @@ -38,7 +49,7 @@ contracts: retrieving focused evidence, returning an answer only with support, a abstaining when no support exists.

- Three evidence-backed examples: focused context holds Recall at 5 while reducing retrieved tokens, answerable questions return supported evidence, and off-topic questions abstain + Three evidence-backed examples: focused context keeps Recall at 5 while reducing returned content, answerable questions return cited support, and unsupported questions explicitly abstain
Each card names its deterministic offline fixture and test scope. The examples are illustrative; they are not customer data or external benchmark results.

@@ -47,21 +58,6 @@ Run `python -m eval.chunking_eval` and `python -m eval.grounded` to reproduce th the former measures evidence retrieval and context size, while the latter measures the answer-versus-abstain decision. -The diagram is the essential path. See [measured context savings](#measured-quality-and-token-efficiency) -for reproducible fixture-level evidence of less returned content at the same tested retrieval -scores, without billing or latency claims. The sections below cover the dashboard, code graph, -local installation, governance controls, and hosted services in detail. - -
- -

- Engraphis Knowledge Graph tab: force-directed entity-relation network -
- Knowledge Graph · run engraphis-dashboard to see it live -

- -
- --- > Open-source users: update regularly for the latest fixes and improvements. diff --git a/docs/images/evidence-backed-agent-examples.png b/docs/images/evidence-backed-agent-examples.png index 7020e21c..07015b0b 100644 Binary files a/docs/images/evidence-backed-agent-examples.png and b/docs/images/evidence-backed-agent-examples.png differ diff --git a/docs/images/evidence-backed-agent-examples.svg b/docs/images/evidence-backed-agent-examples.svg index 4fe66d98..695c48ca 100644 --- a/docs/images/evidence-backed-agent-examples.svg +++ b/docs/images/evidence-backed-agent-examples.svg @@ -1,79 +1,5 @@ - - Evidence-backed examples of Engraphis agent behavior - Three synthetic deterministic fixture examples: focused context reduces returned tokens without a recall regression, an answerable question returns cited support, and an unsupported question abstains. The chart states the exact fixture results and test scope. - - - - - - What an agent gets: tested in deterministic offline fixtures - These are synthetic test inputs and reproducible outcomes, not customer data or external benchmark claims. - - - - - FOCUSED CONTEXT - - - - Find the relevant passage - QUESTION - “What does grounded recall do - when support is low?” - RETURNED EVIDENCE - - “…abstains” - - TESTED: LONG-DOCUMENT FIXTURE - 6 documents · 18 questions - Recall@5 1.000 in both modes - - 808.8 → 218.4 tokens - - - - - - CITED ANSWER - - - Answer when evidence supports it - QUESTION - “Which auth scheme did we - standardise on?” - RETURNED EVIDENCE - - PASETO tokens · source: auth - - TESTED: GROUNDED FIXTURE - 5/5 answerable questions grounded - Answer only with source support - - grounded answer + source - - - - - - SAFE ABSTAIN - - - Refuse unsupported prompts - QUESTION - “How do I bake sourdough - bread?” - RETURNED RESULT - - Abstained: no support - - TESTED: GROUNDED FIXTURE - 5/5 off-topic questions abstained - No source support, no answer - - explicit abstain, no citation - - - - Reproducible scope: - eval.chunking_eval + eval.grounded · 10/10 grounded decisions correct · deterministic offline fixtures + + Three evidence-backed Engraphis agent behaviors + A clear three-card summary of deterministic offline fixtures. Focused context returns 808.8 → 218.4 tokens while retaining Recall at 5 of 1.000. A grounded answer returns support for 5/5 answerable questions grounded. An unsupported question safely abstains for 5/5 off-topic questions abstained. Reproduce with eval.chunking_eval and eval.grounded. +