01General chunks
Keep the right context
Delimiter, length & overlap
Preview the longest answer. Adjust a boundary if it separates a fact from its exception.
How to check the change
The complete supporting fact survives in the returned context.
KNOWLEDGE LAB / PIPELINE REFERENCE
Explore the pipeline, stage by stage.
Compare preparation and retrieval options, read their practical lessons, and study the settings. These choices create a reference checklist.
Start with an application scenario →01 / MAP YOUR PIPELINE
Different sources. One learning path.
Explore a scenario, then build a route around your own documents.
My first chatbot
One file, a few known answers. Read the lesson, or use its recommended settings as a starting point.
Your pipeline
Based on My first chatbot · Recommended settings. Changes here leave the scenario above unchanged.
02Extraction for each source
Choose every kind of content you need to read. Each selected method has its own lesson; apply it only to the matching documents.
After extraction, these sources feed one knowledge base. Its chunk structure, index, and retrieval strategy are shared.
Your selections are saved in this page’s URL. Choosing a scenario only previews its lesson.
Your sources → shared knowledge base
- Local filesReadable text
General chunks Vector + text Hybrid
02 / LEARN & BUILD
Work through every source. Then bring them together.
Your plan has 5 steps. Prepare each source and its extraction methods, then configure the shared knowledge base.
STEP 1 OF 5 / DATA SOURCE
Local files
Start with a small collection whose answers you can verify yourself.
- Choose one current file and give it a recognizable name.
- Write one question and mark the passage that answers it.
- Follow the first-answer lesson; parse the file and record its stable source ID in your ingestion job.
START WITH
harbor-studio-v1.txt
LOOK FOR
Illustrative example · Fictional studio facts · Not a live pipeline runSource record: file name + version + original text
Before moving on
You know which file, version, and passage should answer your first question.
Save this route and its checklist
Includes every selected source, its extraction steps, compatibility notes, and shared tuning guidance. Your page URL also preserves the selected options.
03 / TUNE YOUR CHOSEN ROUTE
One change. A clear comparison.
Build and check your baseline first. These controls follow the choices above.
Build a tiny RAG with one file
02Vector + text
Check the index
Embedding model & source metadata
Record the embedding model and document version; check indexing errors before tuning search.
How to check the change
Known evidence is indexed and traceable. Record actual usage rather than assuming a cost saving.
03Hybrid
Find enough evidence
Weights, Top K & threshold
Try the Top K lesson first. Keep weights and threshold fixed while comparing the returned passages.
How to check the change
Inspect actual passages. Save chosen settings in your search service and answer builder.
04Filters · ranking · citations
Refine and verify
Source scope and evidence quality
Use metadata to keep the intended version in scope. If ranking is the problem, compare reranking separately and measure its added cost and latency.
How to check the change
Check the cited passage, a covered question, and an uncovered question in a fresh conversation.
Changing chunk structure requires reindexing. Compare a different structure in a separate test index. See Technical reference ↗.
04 / VERIFY THE WHOLE PATH
Does the answer hold up?
Connect the knowledge base to your app. For each source, test a known fact and its citation. Then ask a question outside the collection.
After retrieval: check the evidence.
Read the evidence-checking guide →Study how decision models can check relevance and coverage, including the evidence a filter removes.