Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.insitechat.ai/llms.txt

Use this file to discover all available pages before exploring further.

​
Definition

Hybrid search combines two complementary retrieval methods — semantic (vector) search and keyword (BM25) search — and merges their results into a single ranked list. It is the standard retrieval strategy for production-grade RAG systems because pure vector search has known weaknesses that keyword search fills, and vice versa. InsiteChat uses hybrid search with Reciprocal Rank Fusion (RRF) to retrieve content for every chatbot answer.

​
Why one method alone isn’t enough

Good at:
  • Paraphrase matching (“cancel my plan” → “close my subscription”)
  • Cross-lingual matching (“price in rupees” → English pricing page)
  • Conceptual queries that don’t share keywords with the source
Weak at:
  • Proper nouns and brand names (everything looks “similar” to a tech term)
  • Numeric codes, error messages, version numbers
  • Rare industry-specific jargon the embedding model wasn’t trained on
Good at:
  • Exact matches on proper nouns, product names, error codes
  • Numeric and code lookups (“HTTP 429”, “₹1,424/mo”)
  • Rare terminology specific to your industry
Weak at:
  • Paraphrases (the user’s words must overlap with the source)
  • Synonyms (“car” vs “vehicle”)
  • Cross-lingual queries

​
The fix: run both, fuse the results

If you run both methods independently and intelligently merge the rankings, you get the strengths of each and few of the weaknesses.

​
Reciprocal Rank Fusion (RRF)

RRF is a simple, effective algorithm to combine two ranked lists. For each item, its RRF score is: 
RRF_score(item) = 1 / (k + rank_vector) + 1 / (k + rank_keyword)
Where k is a small constant (60 is standard) and rank_x is the item’s position in each ranked list. Items that appear high in either list get a high combined score. Items that appear in both lists get the highest scores of all. The beauty of RRF is that it requires no tuning of weights between vector and keyword scores — only ranks matter — which makes it robust across very different query types. Every chatbot query in InsiteChat runs through this pipeline:
  1. Embed the question with the same model used to embed your content.
  2. Vector search over your indexed chunks for semantic similarity (top 20).
  3. BM25 search over the same chunks for keyword overlap (top 20).
  4. Reciprocal Rank Fusion merges the two lists into a final top-K.
  5. Q&A pair override: if a custom Q&A pair you defined matches with high confidence, it takes priority over both retrieval lists.
  6. LLM generation: the top-K results are formatted into the model’s context window. The LLM generates the answer.
This is why InsiteChat’s answers are accurate on edge cases that trip pure-vector systems — error messages, brand names, pricing in specific currencies, exact phrases from your docs.

​
Hybrid search vs vector-only competitors

Many AI chatbot platforms ship with vector-only retrieval because it is simpler to implement. The trade-off is visible in production:
ScenarioVector-onlyHybrid (InsiteChat)
“What is your refund policy?”✅ Both work✅ Both work
”I got error 429 — what does it mean?”⚠️ May miss exact code match✅ BM25 catches “429"
"How much is the Growth plan in ₹?”⚠️ ₹ symbol may confuse✅ BM25 catches exact symbol
”GST invoice for SaaS plan”⚠️ “GST” may dilute meaning✅ BM25 ranks GST page first
”Cancel subscription” → “Close account” page✅ Vector catches paraphrase✅ Both methods catch it

​
Learn more