Validated paper summary pipeline and runtime options

Day: 2026-05-20
Time: 11:15 to 11:25
Project: Dev
Workspace: WP 2: Operational
Status: In Progress
Priority: HIGH
Assignee: Matías Nehuen Iglesias
Tags: Paperkb, Summary-Pipeline, Jsonl, Validation, Promptflow, Agent-Framework

Description

Validate the paperKB ecosystem and define a bounded architecture for paper-level and corpus-level summaries without overengineering the product loop.

Ran a staged battle-test mindset across the product loop: corpus hygiene, API contract, frontend materialization, export behavior, and failure-mode checks.
Performed a backend smoke-test review that surfaced concrete issues: URL-encoding bugs on paper/chunk endpoints, missing Next.js dependencies preventing frontend startup, duplicate artifact accumulation, missing paper_id propagation in search results, and duplicated export-review outputs.
Evaluated multiple runtime/orchestration options for summary generation, including PromptFlow, Microsoft Agent Framework, OpenAI/Azure batch APIs, LiteLLM, and the llm CLI.
Compared framework usage patterns and converged on a provider-neutral design: keep corpus processing file-based with JSONL inputs/outputs, treat frameworks as optional runtimes, and preserve summaries as derived artifacts owned by the product.
Defined a minimal summary architecture with three tiers, a CLI workflow, JSON schema, selection strategy, and acceptance criteria for a v1 implementation.
Established the MVP gate for summary generation: freeze the artifact contract, enforce idempotency, and make provider integration optional/lazy-loaded.

Clarified that the durable pipeline should remain file-based and JSONL-driven rather than becoming a framework-centric orchestration system.
Narrowed PromptFlow and Microsoft Agent Framework to optional runtime roles, with strategic dependency risk explicitly reduced.
Identified the main product and QA blockers that must be addressed before the paperKB loop can be considered stable.
Produced a concrete implementation direction for summary generation, including backend endpoints, frontend behavior, and test expectations.

Fix URL-encoding handling for paper and chunk endpoints.
Restore or install missing frontend dependencies so the Next.js app can run.
Address duplicate artifact accumulation and export-review duplication.
Ensure search results propagate paper_id correctly.
Implement the stable summary artifact schema and idempotent generation flow.
Build or select the SummaryRunner abstraction for sync and async JSONL batch modes.
Add tests for the summary MVP gate and failure modes.

source_file=2026-05-20.sessions.jsonl, line_number=5, event_count=0, session_id=50efd7dc3fcefc3a1f50a6df12905044dd3fac040f28ec3a87f42532dee480cc
event_ids: []