v1.0.0

Bid Copilot v1.0.0

First stable release. Bid Copilot reads a tender document (招标文件) and automatically generates a structured bid-document outline (投标文件大纲) — the chapter tree a bidder must produce to respond, with every requirement traced back to where it came from in the tender. Built as a demo on Azure OpenAI.

Highlights

9-step understanding pipeline

parse → classify → segment → locate → extract_skeleton → extract_requirements → merge → supplement → finalize

• Regex-first segmentation tuned for Chinese tender numbering; native .docx heading levels preserved.
• Explicit bid-document skeleton extracted from the tender, then normalized and de-duplicated in a two-stage merge (normalize → batched attach) with parallel per-chapter extraction.
• Scoring/commercial requirements extracted item-by-item; technical parameter-level indicators are aggregated into a single "技术参数响应表" entry instead of polluting the outline with one chapter per parameter.
• Engineering guarantees over LLM self-reporting: ref_ids are back-filled by code (not the LLM), and the coverage report is derived from the final tree.
• Requirements that fail to attach during merge are placed by a dedicated supplement step — the LLM only decides where.

Document intake

• Multi-file upload (drag & drop / multi-select) with archive and .ebid package unpacking.
• .docx parsed locally; .doc and scanned PDF go through Azure Content Understanding for structured markdown, with graceful fallback when CU is not configured.

Web UI

• Live step progress over SSE with a vertical-timeline phase log (event-driven sidebar, no polling); running tasks survive page refresh and task switching.
• Outline tree with clickable source badges — hover for provenance, click for a centered modal with location + original quote.
• Coverage panel plus Word export with path-style heading numbering.
• Run sidebar grouped into 运行中 / 已完成, with deletion of completed runs.
• Apple-style minimal UI with a Chinese login page.

Configuration

• Three model tiers — MODEL_MAIN / MODEL_MINI / MODEL_NANO — mapped per pipeline step via MODEL_<STEP>=main|mini|nano.
• Per-step reasoning effort via EFFORT_<STEP>=low|medium|high; parallelism cap via MAX_CONCURRENCY.
• Optional local login gate (ENABLE_LOCAL_AUTH + LOCAL_AUTH_*) for demo deployments.

See the README for Quick Start (Docker / local) and the full configuration reference.