AI ProposalGenerator

This is a working LLM product, one of the AI-integrated systems built at this organization.

1. The knowledge base — an archive of past winning proposal PDFs in a Google Cloud Storage bucket is loaded into a Vertex AI context cache (Gemini 2.0 Flash) on startup: 1-hour TTL, refreshed on demand, with graceful fallback to uncached generation if the cache can't be built.
2. Scoped corpora — GCS folder prefixes act as switchable knowledge bases; the operator picks which slice of the archive the model reasons over.
3. Two operations — a compatibility check ("list matching past proposals and rate the uniqueness of this new request") and proposal generation ("using the style and format of the cached archive, draft a new proposal"). Both prompts live in SQLite via Prisma and are editable in the UI — the operator tunes behaviour without a redeploy.
4. Multi-modal input — the RFP arrives as a PDF, a scanned image, or pasted text. In a region where tender documents are routinely scans, that matters.
5. Handover-shape — dockerized with seeded startup, a complete README, and automatic text-direction handling (dir="auto") for mixed Arabic/English content.

It deliberately isn't RAG. The fashionable architecture in 2025 was a retrieval pipeline — chunking, embeddings, a vector store, tuned similarity thresholds. At this corpus size, that machinery has no payoff. Context caching keeps whole documents intact: style, structure, and the eligibility clauses buried in appendices stay in view. There is nothing to chunk, embed, or tune, and cached input tokens bill at a reduced rate.

Winning grant and tender work is how the organization funds its programmes, and the bottleneck is human attention. The compatibility check answers "have we done something like this before, and how unusual is this request?" before anyone commits days of drafting. The generator then produces a first draft grounded in the organization's own winning style rather than the model's generic idea of a proposal. A person decides what to submit; the tool just gets a first draft onto the page.

The default assumption in 2025 was that grounding a model in documents means RAG. But a retrieval pipeline has real costs: a chunking strategy to design, embeddings to maintain, thresholds to tune — and chunking severs exactly what matters in proposals, where the winning style is a whole-document property and the clause that disqualifies a bid hides in an appendix. Context caching inverts the trade: the entire archive stays intact and visible to every query, at the price of a hard ceiling — the corpus must fit the model's context window, and the cache has a lifecycle to manage (TTL, refresh on archive change, a 1024-token minimum, and a fallback path when creation fails).

The second hard problem is trust. A fluent draft that misreads a mandatory eligibility criterion is dangerous. Grounding generation in the actual archive makes the output checkable against real past work, and a human makes the submission decision.