AI is an asynchronous research & drafting layer that amplifies sourcing agents — it does not make final commercial decisions. Priority use cases, ranked by ROI:
- Supplier research agents — given a company, autonomously gather website, certifications, export history, social, reviews → normalized dossier (
03-product/supplier-intelligence, Phase 2).
- RFQ parsing & classification — turn free-form buyer email into structured specs.
- Discovery agents — crawl source directories/marketplaces and extract candidate supplier records.
- Drafting assistants — intro emails, RFQ covering letters, negotiation notes, QC checklists.
- RAG Q&A over the supplier knowledge base — "find verified PLC suppliers in Shenzhen with ISO + 3 years export history."
- Fit scoring — match dossiers to buyer requests.
| Framework |
Stars* |
License |
Model |
Best for |
Notes |
| LangChain |
143k |
MIT |
Graph/tool orchestration |
Tooling + integrations + RAG |
Huge ecosystem; the "integrations layer" |
| LangGraph |
39k |
MIT |
Stateful graph agents |
Reliable, controllable multi-step agents |
Deterministic control, checkpointing; best for procurement research pipelines |
| CrewAI |
56k |
MIT |
Role-based crews |
Rapid prototyping of "team of agents" |
Opinionated, easy to start; less low-level control |
| AutoGen |
60k |
CC-BY-4.0 |
Multi-agent conversation |
Research/experimentation |
Flexible but heavier to productionize |
| OpenAI Agents SDK |
28k |
MIT |
Lightweight agent/handoffs |
Simple, first-party, production-focused agents |
Minimal, Python/TS; great if already on OpenAI |
| gpt-researcher |
28k |
Apache-2.0 |
Autonomous deep-research |
Long-horizon web research reports |
Built-in web research loop; good scaffold to adapt |
| Eliza (elizaOS) |
18k |
MIT |
Agent framework/OS |
General agents, memory, multi-platform |
TS; heavier; consider if building agent products |
| Microsoft Agents SDK |
1k |
MIT |
Microsoft 365 agent SDK |
MS ecosystem agents |
Only if deep into M365 |
*Star counts fetched 2026-08-07 (see 07-research/github-projects).
Recommendation:
- Primary: LangGraph for agent orchestration (stateful, reliable, testable) + LangChain for tool/LLM integrations and RAG components. Strong for procurement research where deterministic steps matter.
- Alternative (faster/simpler): OpenAI Agents SDK if you standardize on OpenAI and want minimal framework surface.
- Scaffold: adapt gpt-researcher patterns for the deep supplier-research agent.
- CrewAI for quick experiments / demoing a "sourcing team" concept; don't bet the architecture on it.
Documents (dossiers, certs, QC reports, emails, RFQs)
→ chunking (Unstructured / tika for varied file types)
→ embeddings (OpenAI / open-source model)
→ vector store (pgvector early → Qdrant at scale)
→ retrieval: hybrid (BM25 full-text + vector) [Postgres FTS + pgvector]
→ LLM synthesis with citations
- Vector store: start with pgvector on the same Postgres (no new infra on 4 GB). Move to Qdrant (Rust, Apache-2.0, 33k★) when scale/payload filtering demands it. Milvus is heavier; Chroma is fine for dev but not the primary production choice.
- Embeddings: OpenAI
text-embedding-3-* (paid) or a self-hosted open model (e.g., BGE/all-MiniLM, nomic-embed) if cost/offline matters. Budget for OpenAI unless you self-host.
- Hybrid search dramatically improves retrieval on niche industrial part names (e.g., "VFD 15kW" vs. "variable frequency drive").
- Memory / long-term knowledge: consider LangGraph checkpointer for session state; Supermemory/Graphiti (28k★/29k★) if you need a persistent knowledge-graph memory of supplier facts across conversations.
¶ 7.4 Model Context Protocol (MCP) — Agent Tool Standard
- MCP (
modelcontextprotocol/modelcontextprotocol, spec repo 8.8k★; servers reference impl 89k★) is the emerging standard for giving agents tools/data. Official reference servers: filesystem, web search, fetch/HTTP, GitHub, memory, PostgreSQL, browser.
- Why it matters for SupplyShore: instead of bespoke tool code, use MCP servers to expose: web search/browser, Postgres (read supplier data), EspoCRM REST, vector store, company-intelligence APIs. Agents then reuse a standard, swappable tool surface.
- Ecosystem: Activepieces (~400 MCP servers), n8n, Flowise all embrace MCP. Adopting MCP keeps your agent tooling portable across frameworks.
- Recommendation: standardize on MCP for agent tool access; wrap our own APIs (EspoCRM, supplier DB, search) as MCP servers for internal use.
- Human-in-the-loop for anything sent to buyers/suppliers (draft → approve → send).
- Deterministic steps where possible (use LangGraph control flow, not pure LLM autonomy).
- Observability: use Langfuse (32k★) for LLM tracing, evals, prompt management, cost tracking.
- Cost controls: cache embeddings, budget tokens, set max-tool-iterations, cap research depth, prefer cheaper models for extraction vs. synthesis.
- Hallucination guards: require citations (URLs), cross-check facts against structured data, and log source provenance in the dossier.
- Data privacy: do not send PII of buyers/suppliers to untrusted third-party LLMs without consent; prefer providers with data-residency options (see
04-security/security).
Agent orchestration : LangGraph (+ LangChain integrations)
Agent tooling : MCP servers (web search, browser, Postgres, EspoCRM, search)
Research scaffold : gpt-researcher patterns for deep web research
Vector store : pgvector → Qdrant (when scale)
Search : Meilisearch/Typesense (product/supplier search UI) + hybrid RAG
LLM APIs : OpenAI (primary) + open-source fallback (self-host) for cost
Observability : Langfuse
"Supplier Research Agent" — input a supplier name/website → output a structured dossier:
- Crawl website + key pages (Crawlee).
- Extract products, certifications, contact, location (LLM).
- Check company-intelligence + customs APIs (
03-product/apis).
- Score authenticity/fit.
- Write dossier to Postgres + EspoCRM; send notification.
- Log to Langfuse for traceability.
Ship this behind a manual trigger, then wire to W5 (03-product/automation) for automation.