The most cost-effective pattern is usually:

1) Use hybrid retrieval, not a fully agentic “browse everything” loop

Start with:

• BM25 / keyword search for exact terms, IDs, product names, error codes
• Vector search for semantic matches
• Optionally metadata filters first (team, product, date, doc type, permissions)

This is cheaper than making the agent explore many documents or call tools repeatedly.

2) Keep the agent’s job narrow

Use the agent mainly to:

• reformulate the user query
• choose search filters
• rank the top few results
• synthesize the final answer from retrieved snippets

Avoid letting it:

• inspect too many documents
• recursively search in loops
• call the LLM for every chunk unless necessary

3) Add a two-stage retrieval pipeline

A very cost-effective setup is:

1. Fast retrieval
   • Return top 20–50 candidate chunks cheaply
2. Rerank only the top few
   • Use a lightweight reranker or a smaller LLM
3. Generate answer from top 3–5 chunks
   • Keep context small

This usually gives most of the quality benefits at much lower cost than full-agent traversal.

4) Cache aggressively

Cache:

• common queries
• embeddings
• search results
• rerank outputs
• final answers for repeated FAQ-like questions

Internal help docs often have repeated queries, so caching can cut costs a lot.

5) Chunk docs sensibly

For help docs, use:

• medium chunks with headings preserved
• overlap only if needed
• store section titles and breadcrumbs

Bad chunking increases retrieval misses, which causes more agent/tool calls and higher cost.

6) Use an “agentic” system only for harder queries

A cheap routing strategy:

• Simple query → direct retrieval + answer
• Ambiguous query → agent asks clarifying question or expands search
• Complex multi-hop query → agentic workflow

This avoids paying the agent overhead for every request.

7) Prefer a small model for search orchestration

Use a cheaper model for:

• query rewriting
• tool selection
• summarizing snippets

Reserve a larger model only for:

• final synthesis when needed
• complex reasoning across multiple docs

8) Measure cost per resolved query

Track:

• retrieval calls per question
• tokens used in orchestration
• tokens used in final answer
• answer success rate / escalation rate

The most cost-effective design is the one that minimizes cost per successfully answered query, not just raw token usage.

---

Practical recommendation

If you want the best cost/performance balance:

Hybrid search + metadata filters + lightweight reranker + small agent for query rewriting + cache everything.

That usually beats a fully autonomous agentic KB search in both cost and reliability.

If you want, I can also sketch a reference architecture or give you a stack comparison (e.g. Elasticsearch + vector DB + LLM reranker vs pure agentic RAG).