Seocho E2e

SEOCHO e2e (run / sweep)

Drive SEOCHO's YAML-declared e2e: one ontology + one documents folder + questions → index → query → report. No Python; everything is a seocho run or seocho sweep invocation over a run spec. This skill scaffolds the YAML, validates it offline, runs it, and reads the report back.

The YAML is orchestration only. Extraction prompts come from the ontology (node/relationship descriptions) plus optional design specs; the agent pattern is a config key. Keep that separation — do not hand-write prompt strings into the run spec.

0. Resolve the CLI first (do this before any seocho command)

Invoke seocho through uv run against the repo — uv resolves the right environment from pyproject.toml and syncs deps before running, so you never depend on the PATH seocho (often a stale build without run/sweep) or a hardcoded venv path. uv run keeps the current working directory, so relative ontology: / documents: paths in the run spec still resolve.

SEOCHO_REPO=/home/hadry/lab/seocho
if command -v uv >/dev/null 2>&1 && [ -f "$SEOCHO_REPO/pyproject.toml" ]; then
  SEOCHO="uv run --project $SEOCHO_REPO seocho"        # preferred
elif [ -x "$SEOCHO_REPO/.venv/bin/seocho" ]; then
  SEOCHO="$SEOCHO_REPO/.venv/bin/seocho"               # fallback: repo venv
else
  SEOCHO="seocho"                                      # last resort: PATH
fi
$SEOCHO run --help >/dev/null 2>&1 || echo "needs: cd $SEOCHO_REPO && uv sync --extra local --extra ci"

Use $SEOCHO (not a bare seocho) for every command below — e.g. $SEOCHO run seocho.run.yaml --dry-run. If run --help still fails, the repo deps are not synced: run cd $SEOCHO_REPO && uv sync --extra local --extra ci and stop — do not fall back to a stale global binary.

The first uv run in a session may print a one-time dependency-sync line to stderr; that is expected, not an error.

Load the API key the same way the repo does (MARA-first):

set -a && source /home/hadry/lab/seocho/.env && set +a   # exports MARA_API_KEY

Embeddings/LLM default to MARA + local fastembed (bge). Do not introduce OpenAI unless the user asks.

1. Understand the request, then scaffold

Establish three things before writing YAML:

1. Documents — a folder (or file) of .txt/.md/.csv/.json/.jsonl/.pdf. If the user has none, point them at the bundled demo (examples/run/).
2. Ontology — the node/relationship types to extract. If absent, draft a small schema.yaml from the user's domain (see §2), or run seocho init for the interactive builder.
3. Intent knobs — model(s), enforcement mode, agent pattern, questions (§3).

Fastest start (bundled, zero setup) — use this to confirm the toolchain works:

$SEOCHO run /home/hadry/lab/seocho/examples/run/quickstart.yaml --dry-run

seocho run --init writes a fully commented seocho.run.yaml template into the current directory — read it for the authoritative key list before editing.

2. Ontology (schema.yaml)

Separate file the run spec points at. YAML / JSON-LD / TTL. The descriptions are load-bearing — they become the extraction prompt.

name: my-domain
nodes:
  Company:
    description: A business organization        # ← travels into the prompt
    properties:
      name: { type: STRING, constraint: UNIQUE }
  Person:
    properties:
      name: { type: STRING, constraint: UNIQUE }
relationships:
  CEO_OF:
    source: Person
    target: Company
    description: Person leads the company as chief executive

Validate an ontology on its own: $SEOCHO ontology check --schema schema.yaml.

3. Run spec (seocho.run.yaml)

Every key maps to an SDK parameter. Minimal is three keys:

ontology: ./schema.yaml
documents: ./docs/
questions:
  - Who is the CEO of Acme?

Full surface (only add what the user's intent needs):

name: my-run
ontology:
  path: ./schema.yaml
  enforcement: guided          # guided (default) | strict | open  — see §6
documents: { path: ./docs/, recursive: true }
models:
  default: mara/MiniMax-M2.5   # provider/model; MARA-first
  indexing: mara/MiniMax-M2    # optional per-phase override (cheaper indexing)
  query: mara/MiniMax-M2.5
graph: bolt://localhost:7687   # omit → embedded LadybugDB (no server). Or mapping:
graph:
  kind: dozerdb                # neo4j | dozerdb | ladybug
  uri: bolt://localhost:7687
  password: ${NEO4J_PASSWORD}  # secrets via ${ENV}, never inline
vector:                        # optional hybrid search; omit for graph-only
  kind: faiss                  # faiss (in-memory) | lancedb (on-disk)
  embedding: fastembed         # local bge default | provider preset
agent:
  execution_mode: pipeline     # pipeline (default) | agent | supervisor
  routing_policy: balanced     # fast | balanced | thorough
  design: ./agent_designs/x.yaml      # optional AgentDesignSpec (pattern)
indexing:
  design: ./indexing_designs/x.yaml   # optional IndexingDesignSpec (extraction strategy)
query:
  reasoning_mode: true
  repair_budget: 1
  answer_style: concise        # concise | evidence | table
questions:
  - What product does Acme offer?
  - question: Who is the CEO of Acme?
    expect: Jane Park          # recorded in report, not auto-graded

Agent patterns (via agent.design → an AgentDesignSpec; bundled in examples/agent_designs/): reflection_chain, planning_multi_agent, memory_tool_use. Extraction strategy/domain presets live in an IndexingDesignSpec (examples/indexing_designs/, ingestion.extraction_strategy: general|domain|multi_pass). Reference these files rather than re-deriving their internals.

4. Validate, then run

Always dry-run first — it runs the full preflight (ontology loads, documents scanned, API key present, graph reachable, vector deps) with no LLM calls:

$SEOCHO run seocho.run.yaml --dry-run

Fix whatever preflight reports (its messages name the exact fix), then run:

$SEOCHO run seocho.run.yaml

Report lands at runs/<name>-<timestamp>/report.md (+ report.json). Read report.md and summarize: files indexed, nodes/relationships, and per-question answers — flag any empty answer or error.

Useful flags: --only index / --only query (reuse the existing graph), --force (re-index unchanged files), -o DIR, --output-json. Exit codes: 0 ok · 1 runtime/preflight failure · 2 invalid config.

Templating: a *.yaml.j2 run spec is rendered with Jinja2 — pass --var key=value (dotted keys, YAML values) / --vars file.yaml, and --show-rendered to inspect. Secrets stay in ${ENV}, resolved after rendering.

5. Compare variants — seocho sweep

When the user wants to compare configurations (models, enforcement, prompts), use a sweep: one run.yaml.j2 template × N named variants → N isolated runs → one table.

$SEOCHO sweep --init     # writes seocho.sweep.yaml + run.yaml.j2
$SEOCHO sweep examples/run/sweep-enforcement/sweep.yaml --dry-run
$SEOCHO sweep examples/run/sweep-enforcement/sweep.yaml

# seocho.sweep.yaml
template: ./run.yaml.j2
vars: { model: mara/MiniMax-M2.5 }   # shared
variants:
  - name: guided
    vars: { enforcement: guided }
  - name: strict
    vars: { enforcement: strict }

Each variant gets an isolated graph, workspace, and runs/<sweep>-<ts>/<variant>/ dir; the summary table compares files/nodes/answered/empty/errors. Flags: --only-variant NAME, --fail-fast (default keeps going), --var/--vars. Variants run sequentially. Use seocho sweep, not seocho experiment (that one is extraction-only, no query phase).

6. Ontology enforcement (the most-asked knob)

ontology.enforcement is the admission policy for extracted data:

• guided (default) — ontology guides extraction, off-ontology material is written with a warning. Max recall; most QA workloads.
• strict — closed vocabulary: only declared types admitted, chunks with validation errors rejected, no Entity fallback. Trust > recall (audit/regulated). Expect fewer nodes than guided.
• open — admit everything, but stamp out-of-vocabulary nodes/relationships with _out_of_ontology: "true" — the signal for "what should I add to the schema?".

Full reference: https://seocho.blog/sdk/enforcement-modes/

Guardrails

• Resolve $SEOCHO (uv run, §0) and verify run --help works before any run.
• Always --dry-run before a real run; never claim a run succeeded without showing the report or the exit code.
• MARA-first: don't switch to OpenAI embeddings/models unless asked.
• Don't invent prompt text in the run spec — prompts come from the ontology + design specs.
• Secrets go through ${ENV}; never write a literal key into YAML.
• If preflight fails, surface its message verbatim and fix that — don't bypass it.

---
> Source: [tteon/seocho](https://github.com/tteon/seocho) — distributed by [TomeVault](https://tomevault.io).
<!-- tomevault:4.0:skill_md:2026-07-04 -->