xl_summarize — LLM confirmation and input tables

Purpose (non-technical)

Everything through xl_ontology is deterministic—repeatable, auditable, no model calls. xl_summarize is the optional LLM stage for tasks that need human-language judgment on top of the knowledge graph:

• Confirming overall model metadata from cover/intro sheets.
• Building a structured input parameter table for analysts.
• Repairing ambiguous input concept IDs by reading the workbook with a code interpreter.

xl_summarize does not rebuild the DAG or re-detect tables; it reads ontology + annotate artifacts and the cleaned workbook.

---

Three independent capabilities

Capability	LLM?	Output	Use
Overall summary	Yes	final_overall.json	Validate model title, indication, horizon, etc. from intro/cover content
Input table	No	final_input_table.json	List input rows with concept_id, Excel refs, flags—from Pass-1 tags + finalize map
Input parameters repair	Yes	final_input_table_llm.json	Fix rows whose concept is not one of 13 canonical BIM input concepts

You can run any one without the others (orchestrator exposes separate buttons).

---

How the LLM accesses Excel

• Uses OpenAI Responses API with code interpreter.
• The cleaned workbook is attached via container file IDs (not pasted cell dumps)—avoids truncation traps on large models.
• Input-parameters repair issues one batched call to read Sheet!Range contexts and assign concepts or keep.

---

Relationship to xl_structural

Stage	Role
xl_summarize	Targeted BIM workflows: overall confirmation + input inventory
xl_structural (planned)	Broader reasoning over the full knowledge graph

Both are LLM consumers; summarize is product-specific, structural is graph-wide.

---

Outputs

Under data/output/<run_id>/summarize_output/:

• final_overall.json
• final_input_table.json
• final_input_table_llm.json
• Stage folders 01_prep/ … 04_harvest/ when running Overall (debug trail)

---

Technical summary

Entry points (xl_summarize/runner.py)

• run_xl_summarize_overall — prep → upload → respond → harvest
• run_xl_summarize_inputs — deterministic; stages/inputs.py
• run_xl_summarize_input_parameters — stages/input_params_llm.py; consolidated ids inputs.<concept>.<section>[.<subsection>]

Config

• XlSummarizeConfig — config/stages.yaml + env for API keys (core.env_bootstrap)

Upstream reads

• final_tagged_tables.json, final_table_map.json, cleaned_workbook.xlsx, ontology_table_tags.json

Helpers

• client.py, container.py, stages/harvest.py (parse_model_rows, _extract_message_text)

Not part of fast pipeline

• run_fast_pipeline stops after ontology/QC; summarize is manual / orchestrator-triggered.